# LABLOOP · Master Plan v6 — The Cartographer's Path: from program manager to published AI scientist for biotech

**Author:** Peter Ciancaglini · **Date:** 2026-09-13 · **Supersedes:** v5 (2026-09-11). v4 stays as the company map, offers and templates (`AI-Upskilling-Plan-Biotech-v4.md`).
**Horizon:** 40 weeks + week 0, in eight acts. **Load:** 10 h/week, same rhythm as v5.
**What v6 adds, and what it does not touch.** v6 keeps every gate, the thesis, the four papers, the cockpit, the rituals and the prompt pack of v5 exactly as they were. It adds, for every week: a friendly introduction told as a story with an analogy, a map of where the week comes from and where it leads, the intermediate and final goal, a curated learning stack (reading, Stanford Online, Coursera, MIT, YouTube), runnable GitHub repos and Colab notebooks with verified star counts, and a ready-to-build NotebookLM kit. The mentor voice is new; the discipline is not.

> **If you only read one paragraph today:** open Part 6, find your current week, read its "Welcome", build its NotebookLM notebook at the end of Saturday's block, and do the v5 plan inside it. Everything else in this document exists to make that one week easier.

---

## PROLOGUE · A letter from your mentor

Peter,

You have spent 25 years being the person who makes other people's expertise land. You know how to take a room of specialists who disagree, find the real constraint, and ship. That is not nothing; in science it is rare. What you do not have yet is the thing scientists trust before they trust people: **your own measurements, done in a way they can check.**

So here is the deal I am offering you for the next 40 weeks. I will not ask you to become a biologist. I will not ask you to become a machine-learning theorist. I will ask you to become something narrower and more valuable: **the person who turns messy experiments and messy papers into measured, reproducible, decision-ready data, and who can prove it with a DOI.** In biotech that person is scarce, because biologists were trained to run experiments and engineers were trained to build software, and almost nobody was trained to stand in the middle and measure the loop.

Think of yourself as a **cartographer**. For 25 years you coordinated expeditions: budgets, teams, timelines, risks. Other people surveyed the land. Now you are going to learn to survey it yourself. First you learn the instruments (weeks 1-4). Then you build a machine that reads every old explorer's diary and turns it into one map (weeks 5-13). Then you publish the map so other explorers can check it (weeks 14-20). Then a local guide who knows the mountains, a founder-scientist, walks a trail with you and signs the map next to your name (weeks 21-32). Then you get paid to map other people's territories (weeks 27-40). That is the whole story. Every week in this document is one page of it.

Three promises from me, as your mentor:

1. **Nothing will be studied that a build does not need that same week.** Theory without a build is how smart people stay stuck.
2. **Every week has a story, a map and one thing.** You will always know why the week matters, where it goes, and what the minimum is.
3. **Every resource here has been chosen because it helps you understand faster, not because it is famous.** Where a repo has few stars, I say why it is still worth your time (usually: it was written by the scientist whose data you use).

And three things I need from you:

1. Open `TODAY.md` before anything else. Every block.
2. Build the week's NotebookLM notebook at the end of Saturday's buffer block, before the week starts, not during it.
3. When you feel like an impostor, go read the DOIs in `PROGRESS.md`. A DOI is not an opinion.

Let's go.

---

## PROLOGUE 2 · The story of the 40 weeks in eight acts

| Act | Weeks | Story title | The analogy | What you walk out with | Gate |
|---|---|---|---|---|---|
| I | 0-4 | **The apprentice learns the tank** | A new pilot in a flight simulator: first you fly a fake plane, then you replay real flight recorders, then you learn what "normal" looks like so you can spot trouble | A fermentation simulator, 100 real batches understood, statistical alarms, `batchlens v0.1` with a DOI | 4 |
| II | 5-8 | **The reading robot** | Hiring a very fast junior analyst to read 2,000 reports; before trusting them, you grade their homework against an answer key you wrote yourself | FermLit corpus, schema, 50-paper gold set, a measured extraction agent, Technical Report 1 | 8 |
| III | 9-13 | **From sample to census** | Going from a pilot survey of 50 households to a national census, with a second auditor checking your counting | 2,000 papers extracted, kappa, the reporting-completeness finding, FermLit v0.5 on Hugging Face | 13 |
| IV | 14-20 | **Becoming an author** | Moving from writing internal status reports to publishing an audited annual report that outsiders can verify line by line | Paper 1 on bioRxiv (wk 16), Paper 2 on arXiv (wk 20), an endorser, a personal site | 16, 20 |
| V | 21-24 | **The protein detective finds a partner** | A detective who reads fingerprints (protein embeddings) and hands a shortlist of suspects to a forensic lab that can actually test them | `enzymelens`, a ranked candidate list, a co-author agreement, a pre-registered validation plan | 24 |
| VI | 25-32 | **The first client** | The cartographer opens a small office: surveys for hire, fixed price, with published maps as the portfolio | `trialkit`, the Loop Audit offer, first invoice (wk 28), Paper 3 with a co-author (wk 32) | 28, 32 |
| VII | 33-35 | **The lab's memory** | Building the archive and the librarian for a lab: every document findable, every change signed, every answer cited | `labdocs` on eLabFTW, retrieval evaluation, on-prem benchmark | — |
| VIII | 36-40 | **Peer review and the crossroads** | The expedition returns: maps go to the geographic society for review, you give the lecture, and you decide where to settle | Journal + JOSS submissions, a talk, a fund package, a written career decision | 36, 40 |

### The seven threads (how every week connects to the others)

A week never stands alone. It pulls one or more of these threads forward. Each chapter in Part 6-7 names its threads.

| Thread | What it is | Weeks where it moves | Where it ends |
|---|---|---|---|
| **B · Biology literacy** | From "what is an enzyme" to following a lab meeting | 1-24 (Tuesdays) | You can talk to a founder about their organism without faking |
| **M · AI-science method** | From vectors to conformal prediction and hybrid models | 1-22 (Thursdays) | You can defend every number in your papers |
| **BL · batchlens** | Instrument 1: simulator → monitoring → soft sensors → hybrid models | 1-4, 8, 12, 17 | Paper 2 (wk 20), JOSS (wk 36) |
| **FL · FermLit** | Thesis: LLM agents read the fermentation literature, human-validated | 5-7, 9-11, 13-16 | Paper 1 (wk 16), journal (wk 36) |
| **EL · enzymelens** | Protein language models for thermostability, with a partner lab | 16-17 (literacy), 21-24, 30-32 | Paper 3 (wk 32) |
| **$ · Money** | Trialkit, labdocs, Loop Audit, pilots, fund package | 25-40 | First invoice (wk 28), decision (wk 40) |
| **R · Reputation** | Posts, founder papers, ladder, endorsement, talk | Every Sunday and Thursday | Gauge L5 (wk 40) |

### What transfers from your 25 years (and what is genuinely new)

You are not starting from zero. You are starting from a different coordinate system. This table translates it.

| You already know (program management) | The scientist's name for it | Where you use it |
|---|---|---|
| Acceptance criteria before a sprint starts | Pre-registration; "done when" | Week 24 validation plan, every build block |
| Risk register (RAID log) | Limitations section; threats to validity | Every paper's "where it fails" |
| Vendor benchmarking against the incumbent | Baselines | Weeks 8, 12, 17, 22 |
| UAT with a signed test set | Gold standard, held-out test set | Week 6, the most important week of Half 1 |
| Two reviewers sign off a deliverable | Inter-annotator agreement (Cohen's kappa) | Week 10 |
| Audit trail, SOX controls | Provenance; ALCOA+; reproducibility | FermLit provenance sentences, labdocs |
| Status report with RAG colors | Figure with error bars and an interval | Every Sunday post |
| Stakeholder map | Connection ladder, co-authors, endorsers | Part 9 |
| Change control board | Peer review | Weeks 15, 36 |

**What is genuinely new, and where the plan teaches it:** cell biology and metabolism (B1-B8), bioreactor engineering (B9-B12), statistical thinking beyond dashboards (M3-M7, B13-B14), how models learn (M8-M9), uncertainty (M10, M16), and the craft of a scientific paper (B15, M18, weeks 14-20). None of it requires a degree. All of it requires the editor open.

---

## PROLOGUE 3 · The resource rules (so you never waste an evening on a bad source)

1. **Core vs Deeper.** Every chapter splits resources into **▶ Core** (fits inside the Tuesday/Thursday 90 minutes; do these) and **＋ Deeper** (goes into NotebookLM, Saturday sandbox, or only when a concept blocks you). You never owe the Deeper list anything. Finishing a course is never a goal.
2. **Trust signals for code.** Stars were checked on the GitHub API on **2026-09-13** and are rounded (★ 17.5k). Default trust bar: ★ ≥ 1k, active in the last 12 months, or maintained by an institution. **Domain exceptions are marked "niche"**: bioprocess and protein repos rarely reach 1k stars; they are listed because the author wrote the paper or the dataset you use. Archived repos are marked "archived" and are for reading, not depending on.
3. **Any notebook on GitHub opens in Colab.** Replace `https://github.com/` with `https://colab.research.google.com/github/` in the notebook's URL. Example: `https://colab.research.google.com/github/intro-stat-learning/ISLP_labs/blob/main/Ch03-linreg-lab.ipynb`. When a chapter says "Colab", that is what it means. Your RTX 3050 runs the same notebooks locally in `uv` + Jupyter or marimo.
4. **Free first.** Coursera courses can be audited for free (look for "Audit" or "Preview" on the enroll dialog; the certificate is not needed). Stanford Online "free courses" are free. MIT OCW, NPTEL, OpenStax, Khan Academy, LabXchange and the YouTube channels are free. Paid books are marked "(book, paid)" and are always optional.
5. **Verify-at-use.** Links marked **(verify)** were not confirmed live on 2026-09-13; search the title if the link moved. Never cite a paper in your own work that you have not opened.
6. **Language.** Resources are in English. When a Spanish version exists and helps (OpenStax *Biología* 2e exists, for instance), use it for first contact and switch to English for the vocabulary you will publish in.

### The evergreen shelf (optional foundations, never instead of the weekly core)

These are the long courses behind the short weekly units. They are here so you know where the "full version" lives when a unit leaves you hungry. **Rule: only in the Saturday sandbox, only after the week's commitments, and one at a time.**

| Track | Evergreen course | Provider | Why it is on the shelf |
|---|---|---|---|
| Learning itself | *Learning How to Learn* (Oakley & Sejnowski) — coursera.org/learn/learning-how-to-learn | Coursera (McMaster / UCSD) | Chunking, spaced repetition, focused vs diffuse mode. The most ADHD-friendly study course in existence; week 0 uses two of its lessons |
| Biology | *Introduction to Biology: The Secret of Life* (Eric Lander, MIT 7.00x) — edx.org | edX / MIT | The best narrative biology course ever recorded; biochemistry → genetics → recombinant DNA, told as a detective story |
| Biology | MIT OCW 7.016 Introductory Biology (Fall 2018) — ocw.mit.edu/courses/7-016-introductory-biology-fall-2018 | MIT OCW | Lecture videos + problem sets, used in B7 |
| Biology (interactive) | LabXchange — labxchange.org | Harvard (free) | Virtual lab simulations: pipetting, PCR, gel electrophoresis. Seeing the bench helps you read methods sections |
| Bioprocess | *Industrial Biotechnology* — coursera.org/learn/industrial-biotech | Coursera (Univ. of Manchester) | The industrial frame for everything in batchlens and FermLit |
| Bioprocess | NPTEL *Bioreactors* (Prof. G. K. Suraishkumar, IIT Madras) — nptel.ac.in/courses/102106053 | NPTEL | The engineering core of B9-B11, ~10 hours |
| Bioinformatics | *Biology Meets Programming: Bioinformatics for Beginners* — coursera.org/learn/bioinformatics | Coursera (UCSD, Compeau & Pevzner) | Python + DNA, gentle; pairs with Rosalind |
| Statistics | *Introduction to Statistics* (Guenther Walther) — coursera.org/learn/stanford-statistics | Coursera (Stanford) | The statistics a biologist expects you to have, taught by Stanford |
| Statistical learning | *Statistical Learning with Python* (Hastie & Tibshirani) — online.stanford.edu/courses/sohs-ystatslearningp-statistical-learning-python | Stanford Online / edX, free | The video course behind ISLP; YouTube playlist too |
| Machine learning | CS229 *Machine Learning* — cs229.stanford.edu (lectures on the Stanford Online YouTube channel) | Stanford | The math-forward version when ISLP feels too gentle |
| Deep learning | MIT 6.S191 *Introduction to Deep Learning* — introtodeeplearning.com | MIT, free, Colab labs | Fast, modern, with Colab labs; updated every year |
| Bayesian | *Statistical Rethinking 2024* (McElreath) — github.com/rmcelreath/stat_rethinking_2024 (★ 1.8k) | YouTube + GitHub | The kindest Bayesian course; PyMC port in pymc-devs/pymc-resources (★ 2.1k) |
| Bayesian | *Bayesian Statistics* — coursera.org/specializations/bayesian-statistics | Coursera (UC Santa Cruz) | Structured alternative to Rethinking |
| Probabilistic models | *Probabilistic Graphical Models* (Daphne Koller) — coursera.org/learn/probabilistic-graphical-models | Coursera (Stanford) | Heavy; only if causal/graph questions in Paper 3 demand it |
| LLMs | CS336 *Language Modeling from Scratch* — cs336.stanford.edu; CS224N — web.stanford.edu/class/cs224n | Stanford | Evaluation lectures used in M14; everything else is deferred |
| Research craft | CS197 *Computer Science Research* (Michael Bernstein) — cs197.stanford.edu | Stanford | How a research project is actually run, week by week; perfect for a PM turning researcher |
| Knowledge structure | CS520 *Knowledge Graphs* — web.stanford.edu/class/cs520 | Stanford | Schemas, entities, provenance; the grown-up version of FermLit's schema |
| Writing | *Writing in the Sciences* (Kristin Sainani) — coursera.org/learn/sciwrite | Coursera (Stanford) | Used in weeks 14-15, 18, 31 |
| Healthcare AI | *AI in Healthcare* specialization — coursera.org/specializations/ai-healthcare | Coursera (Stanford) | Only if archetype E (clinical) ever becomes Half 3 |
| Drug discovery | *Drug Discovery* — coursera.org/learn/drug-discovery | Coursera (UCSD) | Context for pharma conversations; Half 2 |
| Causality | *A Crash Course in Causality* — coursera.org/learn/crash-course-in-causality | Coursera (Penn) | Alternative to Brady Neal's course for Paper 3 |
| Data management | *Research Data Management and Sharing* — coursera.org/learn/research-data-management-and-sharing | Coursera (UNC / Edinburgh) | FAIR data and data plans; weeks 11, 13, 33 |
| Speaking | *Introduction to Public Speaking* — coursera.org/learn/public-speaking | Coursera (Univ. of Washington) | Week 38 |

### The evergreen channels (subscribe once, search inside them forever)

| Channel | What it is best for | Used in |
|---|---|---|
| **3Blue1Brown** | Visual intuition for linear algebra, calculus, neural networks, probability | M1, M2, M8 |
| **StatQuest with Josh Starmer** | Every statistics and ML concept "clearly explained", with drawings; the best 10-minute unblocker | M3-M5, B13-B14, weeks 2-4, 10 |
| **Steve Brunton** (@Eigensteve) | SVD/PCA, SINDy, Kalman filters, physics-informed ML, control; an engineer's voice | Weeks 2, 12, 13, 17 |
| **Andrej Karpathy** | Neural networks and language models from scratch, in code | M8, M9 |
| **Stanford Online** (@stanfordonline) | CS229, CS224N, CS336, CS25 lectures | M14, evergreen |
| **Crash Course** (Biology, Chemistry, Statistics) | 10-minute narrative introductions to biology and statistics | B1-B6, B13 |
| **Khan Academy** (Biology, Chemistry, MCAT) | Enzymes, respiration, kinetics, genetics, at your own pace with exercises | B1-B7 |
| **Ninja Nerd** | Long, whiteboard-style biochemistry and metabolism lectures | B2, B4 (deeper) |
| **iBiology** (ibiology.org) | Scientists explaining their own discoveries; also career talks | Weeks 17, 21, 24 |
| **LearnChemE** (Univ. of Colorado Boulder) | Short chemical-engineering screencasts: reactors, kinetics, mass balances | B9-B11 |
| **Valence Labs** (@valence_labs) | ML for drug discovery and protein reading groups (M2D2, LoGG) | Weeks 16, 21-23 |
| **ritvikmath** | Data-science concepts on a whiteboard: Bayesian, time series, GPs | M6, M10, M11 |

---

## PROLOGUE 4 · The NotebookLM protocol (one notebook per week, zero decisions)

You asked for a NotebookLM notebook for every week. Here is the protocol so that building it takes 20 minutes and never becomes a project in itself. It lives in the **last 20 minutes of the Saturday buffer block** for next week, so the sacred Sunday hour of Part 0.3 stays unchanged. (If your Google account shows it under a new name, the steps are the same.)

### Naming and structure

- **One notebook per week:** `LABLOOP-W07 · The extraction agent, measured`. Always two digits.
- **One notebook per act (made at the gate):** `LABLOOP-ACT-II · The reading robot`. Its sources are the saved reports (Briefing doc / Study guide) of that act's weekly notebooks plus your gate artifact. This is your compressed memory.
- **One cumulative notebook:** `LABLOOP-THESIS`. Only papers you cite, your own preprints, and `RELATED.md`. It replaces v5's "Bio / ML / Thesis" trio.

### Saturday build (last 20 minutes of the buffer block, for next week)

1. **Create** the notebook with the exact name from the chapter's NotebookLM kit.
2. **Add the chapter itself** as a pasted-text source (copy this week's chapter from this document). This gives NotebookLM the story, the goals and the plan, so every output is about *your* week, not generic.
3. **Add the listed sources** (the kit lists 6-12). What NotebookLM accepts: PDFs, websites (a GitHub repo page works; its README gets indexed), YouTube links (the video needs captions), Google Docs/Slides, pasted text, audio. It does **not** read `.ipynb` files: in Colab use *File → Print → Save as PDF*, or paste the key cells as text.
4. **Add your own material** as it appears during the week: `notes/bio/Bx.md`, `notes/ml/Mx.md`, `QUESTIONS.md`, the week's `LOG.md` lines, figures from your repo. Your own notes are the most valuable sources; they let the Gap Detector (P4) work.
5. **Generate the standard Studio set** below. Do not improvise; the kit gives the customization text.

Free plan limits (as of 2026): 50 sources per notebook and a daily cap on Studio generations. The kits stay at ≤ 15 sources, and the Studio set is 5 outputs, which fits.

### The standard Studio set (same five outputs every week)

| Output | When you use it | Standard customization (paste, then add the chapter's "focus" line) |
|---|---|---|
| **Mind Map** | Open it beside the Tuesday/Thursday reading; click nodes when lost | (no customization; generate after all core sources are loaded) |
| **Audio Overview** (Deep Dive) | Monday walk or commute, 1.25x | `Two hosts: a biotech principal investigator and a senior program manager who is becoming an AI scientist. Explain every concept first with an everyday analogy, then with one analogy from corporate program management, then precisely. Connect everything to the week's build task. End with the three things the listener must be able to explain without notes.` + focus line |
| **Infographic** (style: Scientific or Instructional) | Print or keep as phone wallpaper for the week; post-worthy on gate weeks | `One-page visual of this week's core mechanism as a left-to-right flow with 5-7 boxes, one equation or number per box, and a "where this goes next" arrow at the end.` + focus line |
| **Flashcards** (15) | Saturday, 10 minutes; "Missed it" cards return next Saturday | `Definitions, units, one "why" per concept, and three cards that connect this week to the build task.` |
| **Quiz** (10) | Sunday, before the Gap Detector (P4) | `Start easy, end with two applied questions about my own project and one question that mixes this week with the previous week.` |

**Gate weeks add two outputs:** a **Video Overview** (explainer, 5-8 minutes) that you watch before writing the gate artifact, and a **Slide Deck** you adapt for founder conversations. **Data weeks may add a Data Table** (e.g., extracting parameter values from papers to compare with your own).

### New prompts for NotebookLM (added to the pack as P16-P18 in Part 8)

- **P16 · Visual explainer brief** — turns a concept into an infographic you actually understand.
- **P17 · Connection map** — asks NotebookLM to link this week to earlier weeks' notebooks you pasted in.
- **P18 · Teach-back** — you record a 3-minute voice memo explaining the week, add its transcript as a source, and NotebookLM grades it against the sources.

---

## PROLOGUE 5 · How to read a chapter (the anatomy, identical for all 41)

Each week in Parts 6 and 7 is a **chapter** with the same nine blocks, always in the same order, so your eyes learn where things are:

1. **Header line** — act, threads, next gate.
2. **Welcome** — the friendly introduction, told as a story with one central analogy.
3. **Where this week sits** — ← what you bring from earlier weeks · → what this week enables later · ⟳ where it comes back.
4. **Goals** — 🎯 this week (intermediate) · 🏁 next milestone (gate) · 🌟 final (week 40).
5. **The plan** — the v5 weekly brief, unchanged for weeks 0-20; expanded from the v5 table for weeks 21-40 without changing gates or outputs.
6. **Learn it** — ▶ Core and ＋ Deeper resources.
7. **Run it** — repos and Colab notebooks with ★ and a concrete 20-40 minute exercise.
8. **NotebookLM kit** — notebook name, sources to load, focus line for the Studio set, one custom question.
9. **Mentor check** — the one sentence you should be able to say at the end of the week.

---

## PART 0 · How to use this document (read this every Monday for the first month)

### 0.1 The honest answer to your question

You asked how a program manager with 25 years and strong AI fluency can walk into a biotech company and sell something he built, in 20 weeks, like a pro. He cannot, and the v4 plan was wrong to imply it. Here is what is true:

1. **You will not sell biology. You will sell measurement.** Biotech founders know their biology better than you ever will. What they do not have is someone who can turn their experiments into structured data, models and decisions, and who can prove it with a paper. That person is rare and is paid accordingly.
2. **Credibility in science has only one currency: things other scientists can check.** A repo with a DOI, a dataset other people download, a preprint with a method they can reproduce, a co-authored paper with a scientist who has a lab. Not talks, not posts, not certificates. Posts only amplify what the currency already bought.
3. **The bridge from "outsider with a tool" to "peer" is a co-author.** Most GridX founders are CONICET scientists. When one of them puts their name next to yours on a preprint, you are no longer a consultant pitching them; you are a collaborator they introduce to others. Everything in this plan is arranged so that this happens between week 16 and week 32.
4. **Twenty weeks is enough for the first two rungs, not for the pro.** At week 20 you will have one bioRxiv preprint (a dataset and a method), one endorsed arXiv preprint (a modeling result), three instruments with numbers, and a reading level in biology and machine learning that lets you follow a lab meeting. That opens the employee track at scientific-data companies and makes founders answer your messages. The paid consulting starts in Half 2, after the third paper is in motion, around week 28-32.
5. **A plan cannot guarantee a discovery. It can guarantee the artifacts that make one likely.** Three of the four papers in this plan are of a kind that reliable work produces (a dataset, a benchmark, a reproducible modeling study). The fourth (the enzyme paper, Half 2) is the one with a real chance of a finding, and it is designed with a founder's lab as the place where the finding gets tested.

### 0.2 The cockpit: five files, nothing else

Create a private repo `cockpit` on day 0. It holds five files. You open it at the start of every block and close it at the end. Nothing about the plan lives in your head.

| File | What goes in it | When |
|---|---|---|
| `TODAY.md` | One line: the next physical action ("open `sim.py`, add the substrate equation"). Never a goal, always an action | End of every block, for the next one |
| `PROGRESS.md` | The 40-week checklist from Part 6 and 7, copied verbatim, with `[ ]` boxes | Tick boxes on Sunday |
| `LOG.md` | One line per block: date, block type, what happened, one number if any | End of every block, 60 seconds |
| `PARKING.md` | Every shiny idea, tool, paper, project, product that is not the current task. Write it, close it, move on | Whenever the idea appears |
| `QUESTIONS.md` | Every "I don't get this" while studying. These feed the prompts on Thursday and Sunday | While studying |

### 0.3 The rituals (this is the ADHD armor)

**Start ritual (2 minutes):** open `cockpit/TODAY.md`, read the one line, start a 90-minute timer (or two 40-minute halves with a 10-minute walk), one window, one repo, phone in another room. If you freeze, use Prompt P12.

**Stop ritual (3 minutes):** write tomorrow's one line in `TODAY.md`, one line in `LOG.md`, commit, close everything. The block ends when the timer ends, even mid-thought; the mid-thought goes in `TODAY.md`.

**The rule of one:** one block, one task, one repo, one file open in the editor when possible. Theory or code, never both in a block.

**The parking lot rule:** every idea that is not today's task goes to `PARKING.md` in under 20 seconds. You will review it on Saturday. Most of them die there; the good ones become Half 2.

**Saturday sandbox:** the Saturday block is the only place where hyperfocus on a parked idea is allowed, and only if the week's Monday-Friday blocks are done. It is a reward, and it is bounded by the timer.

**Sunday hour (sacred):** 20 min build log, 5 min tick boxes, 20 min Prompt P4 (gap detector), 15 min plan the week's six `TODAY.md` lines in advance (write all six now; during the week you only read).

**Lost week rule:** a lost week is expected. Two in a row is the only forbidden thing. After a lost week, run Prompt P14 on Sunday and cut the next week to the "if you only do one thing" line.

### 0.4 Weekly shape (never changes for 40 weeks)

| Day | Block | Content |
|---|---|---|
| Mon | Build 1 | Thesis or instrument code |
| Tue | Study Bio | Biology track (Part 4.1), editor open, notes into `QUESTIONS.md` |
| Wed | Build 2 | Thesis or instrument code. Last 10 min: Wednesday note (Part 9) |
| Thu | Study AI | AI-science track (Part 4.2). Last 30 min: one founder paper (Part 4.4) |
| Fri | Build 3 | Thesis or instrument code |
| Sat | Buffer | Whatever slipped; if nothing slipped, sandbox |
| Sun | Publish | The sacred hour |

Every week in Parts 6 and 7 is a chapter (Prologue 5): a story-based welcome, where the week sits, goals, the v5 plan (goal, Tuesday and Thursday material, three build tasks with "done when", Sunday output, "only one thing"), Core and Deeper resources, runnable repos, a NotebookLM kit and a mentor check. **v6 addition:** the last 20 minutes of the Saturday buffer build next week's NotebookLM notebook (Prologue 4).

---

## PART 1 · The strategy in one picture

```mermaid
flowchart TB
  subgraph H1 [Half 1 · weeks 0-20 · the scientific profile]
    S[Study: Bio track + AI-science track + domain track]
    I1[Instrument 1 batchlens: simulator, MSPC, soft sensor, deviations]
    T1[Thesis project FermLit: LLM agents read 2,000 fermentation papers, human-validated gold set]
    P1[Paper 1 · bioRxiv wk16: FermLit dataset + method + what papers fail to report]
    P2[Paper 2 · arXiv wk20 endorsed: hybrid soft sensors under data scarcity]
    S --> I1 --> P2
    S --> T1 --> P1
    T1 --> P2
  end
  subgraph H2 [Half 2 · weeks 21-40 · depth, co-author, money]
    P3[Paper 3 · enzyme thermostability with protein LMs, founder co-author, wet-lab validation]
    J[Journal submission of Paper 1 or 2]
    I23[Instruments 2 and 3: trialkit, labdocs]
    M[Loop Audit → first invoice wk 28-32 → pilots → retainer or job]
    P1 --> P3
    P2 --> J
    P3 --> M
    I23 --> M
  end
  R[Reputation gauge L0→L6] -.-> H1
  R -.-> H2
```

Read it like this: study feeds two things, an instrument and a thesis project. The instrument produces a modeling paper. The thesis project produces a dataset paper. Both papers plus the instrument make founders answer. One founder becomes a co-author on Paper 3. Paper 3 plus the instruments become the thing you get paid for. The reputation gauge (Part 2) tells you which rung you are on at any moment.

---

## PART 2 · The reputation gauge (what "reputation" means, measurably)

| Level | Name | Evidence that you are here | Unlocks | Target week |
|---|---|---|---|---|
| L0 | Private | Code on your disk | Nothing | Today |
| L1 | Public and reproducible | Public repo, README with a number, `make report` regenerates it, `--fake` mode | Engineers take you seriously | 4 |
| L2 | Citable | Zenodo DOI on a release; Hugging Face dataset card | Scientists can cite you | 8 |
| L3 | Preprinted | bioRxiv preprint (no endorsement needed) with a method others can reproduce | Founders answer messages; scientific-data companies interview you | 16 |
| L4 | Endorsed and co-authored | arXiv preprint with an endorser; a scientist's name next to yours | You are a peer, not a vendor | 20 (endorsed), 28 (co-author) |
| L5 | Reviewed and invited | Journal or JOSS acceptance; a talk at a biotech event | High-ticket audits; fractional roles | 36-40 |
| L6 | Used | Someone else's paper cites FermLit; a lab uses batchlens | The market comes to you | Month 12+ |

The gauge is the only progress metric that matters. Every Sunday you write the current level in `PROGRESS.md`.

---

## PART 3 · The scientific thesis (explained twice: like a kid, then like a reviewer)

### 3.1 Like a kid

Thousands of scientists have grown microbes in tanks to make useful things: medicines, pigments, enzymes, plastics. Each one wrote what they did in a PDF, each in a different way. Nobody has put all of that into one big table. You are going to build a robot that reads those PDFs and fills the table, and you are going to check the robot's homework by hand so everybody trusts the table. Then you will be the first person who can say what the table shows: what most papers forget to write down, and what a computer can and cannot predict about a tank from the numbers people bother to report. Along the way you build a tool that watches a tank in real time and says "something is going wrong" hours before a human would notice. Then you use the table and the tool to help a real scientist find better enzymes, and you write that up together.

### 3.2 Like a reviewer

**Thesis statement.** *Published bioprocess literature is an underused, unstructured dataset. Large language model agents, validated against a human-annotated gold standard with published inter-annotator agreement, can convert it into a structured resource that (a) reveals systematic gaps in what fermentation papers report, and (b) enables data-scarce process modeling to be benchmarked against mechanistic and hybrid baselines. The same rigor, applied to protein language models, yields testable candidates for industrially relevant enzymes.*

**The four papers**

| # | Working title | Type | Venue | Week | What makes it real science |
|---|---|---|---|---|---|
| 1 | *FermLit: an LLM-agent-curated, human-validated dataset of N published fermentation experiments, and what the literature fails to report* | Dataset + method + descriptive analysis | bioRxiv (no endorsement needed), later Scientific Data or a bioprocess journal | 16 | Novel dataset; gold set with two annotators and Cohen's kappa; per-field precision/recall; a reporting-completeness finding nobody has quantified for fermentation at scale |
| 2 | *Hybrid mechanistic-ML soft sensors under data scarcity: a reproducible study on an open industrial fermentation benchmark* | Modeling study | arXiv cs.LG or q-bio.QM (endorsed), later Computers & Chemical Engineering or Biotechnology and Bioengineering | 20 | Learning curves (5, 10, 20, 50 batches), mechanistic vs pure ML vs hybrid, transfer to faulty batches, uncertainty via conformal intervals, all reproducible in one command |
| 3 | *Protein language model embeddings predict thermostability of industrially relevant enzymes from extremophiles: candidates and validation* | Applied ML + wet-lab validation | bioRxiv then a journal, with a founder-scientist co-author | 28-32 | ESM-2 embeddings + small heads on FLIP/Meltome; extremophile focus; a ranked candidate list a partner lab tests. This is the paper with a chance of a discovery |
| 4 | *batchlens / FermLit software* | Software paper | JOSS | 36 | Peer-reviewed software; cheap credibility for the instruments |

**Why this thesis and not another**

- It uses what you already have (agents, RAG, orchestration, evals) as the *method*, so the science is where you are strong, and the biology is where you are learning.
- It needs no wet lab until Paper 3, where the wet lab belongs to the co-author.
- Every piece is public data: Europe PMC open-access full texts, IndPenSim, FLIP and Meltome.
- It fits 6 GB of GPU: extraction runs on APIs and local 7B models; soft sensors train on CPU; ESM-2 650M embeddings run in fp16 on the 3050.
- It produces things founders in archetype A (25 companies in the GridX list) will want to read, and things scientific-data companies (DataHow, Aizon, TetraScience) hire for.

**Risks and what you do about them**

| Risk | Mitigation |
|---|---|
| Someone already published a fermentation literature dataset | Week 5, Prompt P8 + Semantic Scholar search. If it exists, you extend it (different organisms, newer years, richer schema) and cite it. Extension is still a paper |
| Extraction accuracy too low to trust | The gold set is built first (week 6), the pipeline is tuned against it, and the paper reports whatever the number is. A well-measured 78% is publishable; an unmeasured 95% is not |
| No endorser by week 20 | Three parallel routes (Part 9.5). Fallback: Paper 2 also goes to bioRxiv, and arXiv follows when the endorser appears. The gate is "preprint submitted", the endorsed arXiv is the target |
| Paper 3 partner never materializes | Paper 3 is still publishable as a computational study with public data. The partner adds validation and the co-author line, not the existence of the paper |
| You cannot judge if the science is sound | Two review lines: Prompt P10 (reviewer 2) every draft, and one human scientist reads each paper before submission (the co-author, the endorser, or a paid 2-hour review from a postdoc, ~USD 100) |

---

## PART 4 · Curriculum from zero (four tracks, all free, exact units)

Rules: every item lists what you produce and when it is "done". You never watch without the editor open. `QUESTIONS.md` is open in every study block. Nothing is studied that no week in Part 6 asks for. **v6:** the tables below are the index of the study units; each chapter in Parts 6-7 adds the full Core/Deeper resource list, repos and NotebookLM kit for its unit.

### 4.1 Track B · Biology from zero (Tuesdays, ~30 h in Half 1)

| Unit | Weeks | Material (exact) | What you produce | Done when |
|---|---|---|---|---|
| B1 · Molecules of life and enzymes | 1 | OpenStax *Biology 2e* ch. 3 (biological macromolecules) and ch. 6.5 (enzymes) — openstax.org/details/books/biology-2e. Crash Course Biology "Biological Molecules" (YouTube) | 1 page of notes in `notes/bio/B1.md` + 5 questions | You can say what a protein, an enzyme and a substrate are without looking |
| B2 · The cell and energy | 2 | *Biology 2e* ch. 4 (cell structure, skim), ch. 7 (cellular respiration), ch. 5.1-5.2 (membranes) | Notes + a drawing of a cell making ATP | You can explain why a microbe needs oxygen and sugar in a tank |
| B3 · Microbial growth | 3 | OpenStax *Microbiology* ch. 9 "Microbial Growth" — openstax.org/details/books/microbiology | Notes + the growth curve drawn by hand with lag, log, stationary, death | You can connect the log phase to the Monod equation in your simulator |
| B4 · Microbial metabolism and fermentation | 4 | *Microbiology* ch. 8 "Microbial Metabolism" (8.1-8.4) | Notes; one paragraph "what is a fermentation, really" | You can explain aerobic vs anaerobic to a friend |
| B5 · DNA to protein | 5 | *Biology 2e* ch. 14 (DNA), 15 (genes and proteins); Crash Course "DNA, Hot Pockets & the Longest Word Ever" | Notes; Rosalind problems 1-4 (rosalind.info) | You can read "the strain overexpresses gene X" and know what it means |
| B6 · Genetic engineering and industrial strains | 6 | *Microbiology* ch. 12 "Modern Applications of Microbial Genetics"; *Biology 2e* ch. 17.1 | Notes; Rosalind 5-8 | You understand what "engineered *Yarrowia*" or "recombinant *E. coli*" means in a FermLit paper |
| B7 · Protein structure and enzyme kinetics | 7 | MIT OCW 7.016 Introductory Biology (Fall 2018) lectures on protein structure and enzymes — ocw.mit.edu (search 7.016); Khan Academy "Enzyme kinetics" | Michaelis-Menten fitted in Python to a public dataset | You can fit Km and Vmax and say what they mean |
| B8 · Industrial biotech overview and review | 8 | *Microbiology* ch. 12.5 and any open "industrial microbiology" overview; re-read your B1-B7 notes | A 2-page "biology for my thesis" summary | Prompt P3 quiz score ≥ 4/5 |
| B9 · Bioreactors I: modes and balances | 9 | NPTEL "Bioreactors" (IIT Madras, Prof. G. K. Suraishkumar), lectures 1-3 — nptel.ac.in (verify current playlist); MIT OCW 10.37 lecture notes on bioreactors as backup | Mass balances for batch, fed-batch, continuous in a notebook | Your simulator's equations match the lecture's |
| B10 · Bioreactors II: kinetics and yields | 10 | NPTEL lectures 4-6 (Monod, yields, maintenance) | Yield coefficients computed on IndPenSim | You can explain Yx/s and Yp/x |
| B11 · Bioreactors III: oxygen transfer and scale-up | 11 | NPTEL lectures on kLa, mixing, scale-up | kLa estimated from IndPenSim DO data | You can say why scale-up fails |
| B12 · Downstream, PAT and QbD | 12 | FDA PAT guidance (2004) — fda.gov; ICH Q8(R2) sections 1-2 — ich.org | Glossary: CPP, CQA, design space, golden batch, in your words | You can hold a conversation with a pharma QA manager |
| B13 · Statistics for biologists I | 13 | *Modern Statistics for Modern Biology* (Holmes & Huber) ch. 1-2 — huber.embl.de/msmb | Two exercises redone in Python | You know why biologists distrust p-values from n=3 |
| B14 · Statistics for biologists II | 14 | MSMB ch. 6 (testing, multiple testing) | Multiple-testing correction applied to a FermLit comparison | You can defend an effect in Paper 1 |
| B15 · How papers are built | 15 | Read 3 fermentation papers from FermLit with Prompt P11; note the IMRaD structure | Structured digests in `notes/papers/` | You can find methods, results and limitations in 10 minutes |
| B16 · Protein language models | 16 | Lin et al. 2023, "Evolutionary-scale prediction of atomic-level protein structure with a language model" (ESM-2), read the abstract, intro and figures; MIT 6.874 protein lecture (Kellis, YouTube "Deep Learning in Life Sciences 2021") | ESM-2 650M running on the 3050, embedding 100 sequences | You can explain what an embedding of a protein is |
| B17 · Enzyme stability and engineering | 17 | FLIP benchmark paper (Dallago et al. 2021, "FLIP: Benchmark tasks in fitness landscape inference for proteins"); Meltome atlas paper (Jarzab et al. 2020) | Meltome split loaded; baseline on 1,000 sequences | You know what "thermostability" is measured as |
| B18 · Extremophiles | 18 | Two founder papers (Puna Bio, Antarka, APEXzymes authors) via Prompt P11 | Outreach lines drafted | You can name the organisms and enzymes they work with |
| B19 · Omics literacy | 19 | scanpy PBMC 3k tutorial — scanpy.readthedocs.io | Notebook run end to end | You can follow a single-cell talk |
| B20 · Review and mock lab meeting | 20 | Re-read all notes; Prompt P3 with "lab meeting" persona | A 10-minute self-recorded explanation of Paper 1 for biologists | You can present without notes |

### 4.2 Track M · AI science from the basics (Thursdays, ~30 h in Half 1)

| Unit | Weeks | Material (exact) | What you produce | Done when |
|---|---|---|---|---|
| M1 · Linear algebra, intuitively | 1 | 3Blue1Brown "Essence of Linear Algebra" ch. 1-7 (YouTube, 1.25x) | Notes; numpy re-doing 3 examples | You can say what a matrix does to a vector |
| M2 · Calculus and the neural network idea | 2 | 3Blue1Brown "Essence of Calculus" ch. 1-4 + "Neural Networks" ch. 1-4 | Notes | You can explain a gradient with a hill |
| M3 · Statistical learning I | 3 | *An Introduction to Statistical Learning with Python* (ISLP) ch. 2 and 3 — statlearning.com (free PDF + lecture videos) | Lab 3 redone on IndPenSim variables | You know bias-variance and can fit and read a regression |
| M4 · Validation done right | 4 | ISLP ch. 5 (cross-validation, bootstrap) | Batch-wise CV implemented in batchlens | You can explain leakage across batches |
| M5 · Regularization and trees | 5 | ISLP ch. 6.1-6.2 and ch. 8 (trees, boosting) | LightGBM soft sensor with tuned regularization | You can explain why boosting wins on tabular data |
| M6 · Probability, the useful part | 6 | Murphy, *Probabilistic Machine Learning: An Introduction* ch. 2 (probability) and 4.1-4.3 (statistics) — probml.github.io (free PDF); or Statistical Rethinking 2024 lectures 1-2 (github.com/rmcelreath/stat_rethinking_2024) | Notes; a Bayesian update by hand | You can explain a posterior with coins |
| M7 · Bayesian models | 7 | Statistical Rethinking 2024 lectures 3-4 | A simple Bayesian regression in PyMC or numpyro on titer data | You can read a credible interval |
| M8 · Backpropagation by hand | 8 | Karpathy, "The spelled-out intro to neural networks and backpropagation: building micrograd" — karpathy.ai/zero-to-hero.html | micrograd rewritten from memory | You can derive a gradient of a tiny network |
| M9 · Language models from scratch | 9 | Karpathy, "Building makemore" (part 1) | Bigram model on organism names from FermLit | You know what a language model is, in 200 lines |
| M10 · Gaussian processes | 10 | Rasmussen & Williams, *GPML* ch. 1-2 — gaussianprocess.org/gpml (free); scikit-learn GP examples | GP soft sensor with uncertainty bands | You can explain a kernel as "similarity" |
| M11 · Bayesian optimization | 11 | Frazier, "A Tutorial on Bayesian Optimization" — arxiv.org/abs/1807.02811; BoTorch tutorials — botorch.org | BO on the simulator's feed profile | You can explain expected improvement |
| M12 · Data-driven dynamics | 12 | Brunton & Kutz, *Data-Driven Science and Engineering*, ch. 7 (SINDy) — databookuw.com (free PDF + YouTube "Steve Brunton") | SINDy recovers Monod from simulator data | You can say what "system identification" is |
| M13 · Hybrid and scientific ML | 13 | von Stosch et al. 2014, "Hybrid semi-parametric modeling in process systems engineering: past, present and future"; Chen et al. 2018, "Neural Ordinary Differential Equations" (abstract + fig. 1 only) | Hybrid model v0: Monod + GBM residual | You can draw the hybrid architecture |
| M14 · Evaluation and judges | 14 | Stanford CS336 lecture 12 (evaluation) — cs336.stanford.edu; CS224N benchmarking lecture (judge biases) | Calibrated judge for FermLit fields | You can explain judge-human agreement |
| M15 · Information extraction, measured | 15 | Jurafsky & Martin, *Speech and Language Processing* (3rd ed. draft, free) chapter on information extraction; Cohen's kappa (any statistics reference) | Per-field precision/recall, kappa between annotators | Paper 1's evaluation section is written |
| M16 · Uncertainty you can defend | 16 | Angelopoulos & Bates, "A Gentle Introduction to Conformal Prediction" — arxiv.org/abs/2107.07511, sections 1-2 | Conformal intervals on the soft sensor | Paper 2 reports coverage |
| M17 · Deep learning for proteins | 17 | MIT 6.874 protein lecture; Hugging Face ESM model card (facebook/esm2_t33_650M_UR50D) | Embedding + ridge head on Meltome | Baseline number exists |
| M18 · Research craft | 18 | Simon Peyton Jones, "How to write a great research paper" (YouTube, 1 h); NeurIPS paper checklist (neurips.cc) | Paper 2 rewritten with the checklist | Reviewer-2 prompt finds no unsupported claim |
| M19 · Effect sizes and honest reporting | 19 | MSMB ch. on reporting; bootstrap intervals (ISLP ch. 5.2) | Every number in Papers 1-2 has an interval | Yes/no |
| M20 · Review and mock interview | 20 | Prompt P3, "staff ML scientist at a bioprocess company" persona | Recorded 10-minute explanation of Paper 2 | You can present without notes |

Deferred to Half 2 (Part 7): Karpathy videos 2-8 as needed, fast.ai if you want image models for trialkit, causal inference (Brady Neal's free course) for Paper 3's analysis, CS336 lectures 1-2 and 10 only if labdocs needs them.

### 4.3 Track D · Domain and craft (inside build blocks, no separate hours)

| Item | Link | Used in |
|---|---|---|
| IndPenSim dataset and papers (Goldrick et al. 2015, *J. Biotechnol.*; Goldrick et al. 2019, *Comput. Chem. Eng.*) | industrialpenicillinsimulation.com (verify) | Weeks 2-8, Paper 2 |
| BioModels (curated mechanistic models, SBML) | ebi.ac.uk/biomodels | Week 12 (compare your Monod to a curated model) |
| Europe PMC REST API (open-access full text) | europepmc.org/RestfulWebService | FermLit corpus, week 5 |
| PMC Open Access subset | ncbi.nlm.nih.gov/pmc/tools/openftlist | FermLit corpus |
| Semantic Scholar API | api.semanticscholar.org | Related-work checks, Prompt P8 |
| NIST/SEMATECH e-Handbook ch. 6 (process monitoring) | itl.nist.gov/div898/handbook | Weeks 3-4 |
| Hugging Face datasets docs (dataset cards) | huggingface.co/docs/datasets | Week 13 |
| Zenodo GitHub integration | zenodo.org | Weeks 4, 8, 13, 17 |
| bioRxiv submission guide | biorxiv.org/submit-a-manuscript | Week 16 |
| arXiv endorsement rules | info.arxiv.org/help/endorsement.html | Weeks 12-20 |
| JOSS author guide | joss.theoj.org | Half 2 |
| eLabFTW docs | doc.elabftw.net | Half 2, labdocs |

### 4.4 Track F · Founder papers (Thursday, last 30 minutes)

One paper per week by a founder or scientific lead of a GridX company (archetypes A, C, D first). Use Prompt P11. Output: one digest in `notes/papers/` and one outreach line saved for later. By week 20 you have 20 digests and 20 lines. This is the cheapest, highest-yield thing in the plan: nobody else who writes to them has read their paper.

---

## PART 5 · Repos and materials to prepare (day 0 to week 2)

### 5.1 Repositories

| Repo | Visibility | Purpose | Skeleton |
|---|---|---|---|
| `cockpit` | Private | The five files of Part 0.2 | `TODAY.md`, `PROGRESS.md`, `LOG.md`, `PARKING.md`, `QUESTIONS.md` |
| `labloop-notes` | Public | Study notes, paper digests, weekly build logs. Learning in public is content | `notes/bio/`, `notes/ml/`, `notes/papers/`, `logs/` |
| `batchlens` | Public | Instrument 1 and the code of Paper 2 | `sim/`, `data/` (download script, never the data), `mspc/`, `softsensor/`, `hybrid/`, `report/`, `tests/`, `Makefile`, `CITATION.cff`, `README.md` |
| `fermlit` | Public | Thesis project and the code of Paper 1 | `corpus/` (queries, PMC ids), `schema/`, `annotate/` (guidelines, gold set), `extract/` (agents), `eval/`, `analysis/`, `dataset/` (HF card), `paper/` |
| `paper-2-hybrid-softsensor` | Public at submission | LaTeX or Quarto source, figures regenerated by `make` | `paper/`, `figs/`, `Makefile` |
| `enzymelens` (Half 2) | Public | Paper 3 code | later |

### 5.2 README skeleton (same for every instrument repo)

```
# <name>

<One sentence: what it measures, on which public data.>

**Headline number:** <the number, with N and interval>. Everything below depends on that.

## 60-second quickstart
    pip install -e .
    make report FAKE=1      # full pipeline, no data, no keys
    make report             # downloads public data, regenerates every figure

## What it does / What it does not do
## Data (public, with download script and checksums)
## Reproduce the paper: `make paper`
## Where it fails (honest section)
## Cite: CITATION.cff, Zenodo DOI badge
```

### 5.3 Local environment (day 0, one block)

- Python 3.11 via `uv`; one virtual env per repo.
- `ollama pull qwen2.5:7b-instruct-q4_K_M`, one vision model (`gemma3:4b` or `qwen2.5vl:3b`), one embedding model (`bge-m3`). Confirm each loads on the 3050.
- API keys in a `.env` never committed; `--fake` adapters in every pipeline (the v3 idea).
- Debian box: Docker with Mosquitto, InfluxDB, Grafana (week 1), eLabFTW (Half 2).
- NotebookLM (**changed in v6**): one notebook per week (`LABLOOP-Wnn`), one per act (`LABLOOP-ACT-x`, made at the gate) and one cumulative `LABLOOP-THESIS`. Weekly notebooks stay at ≤ 15 sources. Protocol in Prologue 4; calendar in Appendix V6-A.

### 5.4 Annotation kit for FermLit (week 6)

- `annotate/GUIDELINES.md` written with Prompt P7: one page per field, with positive and negative examples.
- `annotate/gold/` : 50 papers, each a JSON with the fields and the exact sentence that supports each value (provenance is what reviewers ask for).
- A second annotator for 30 of the 50 (a scientist friend, a founder, or a paid biology student, ~USD 100) so you can report Cohen's kappa.

---

## PART 6 · Weeks 0-20 as chapters (the scientific profile)

Format per chapter: see Prologue 5. The **plan** block inside each chapter is the v5 brief verbatim. G = GitHub, L-Sun = LinkedIn build log, L-Wed = Wednesday note, M = Medium, Z = Zenodo DOI, HF = Hugging Face, X = preprint.

---

# ACT I · The apprentice learns the tank (weeks 0-4)

> **Act I story.** A new pilot does not start in a real cockpit over the Andes. They start in a simulator, where crashing is free. Then they replay real flight recorders from other pilots. Then they learn what a normal flight looks like, statistically, so they can tell when something is going wrong before the passengers notice. Finally they get their first licence. Your plane is a fermentation tank; your flight recorders are the 100 IndPenSim batches; your licence is `batchlens v0.1` with a DOI.

---

### Week 0 · Cockpit and positioning

**Act I · Threads: R (reputation), all repos · Next gate: 4**

**Welcome.** Before a long expedition, good expedition leaders spend a whole day on something that looks boring: packing. Every item has a pocket, every pocket has a label, and the map is on top. They do this because on day 20, exhausted and cold, they will not have the willpower to search for the compass. Week 0 is packing day. You are not learning biology yet and you are not writing a simulator yet. You are building the cockpit that will carry your attention for 40 weeks: five files, three repos, a local model that answers, an ORCID, a headline that tells the world you have started. It feels small. It is the most leveraged day of the plan, because a twice-exceptional brain does not fail for lack of intelligence; it fails when the next action is not obvious. After today, the next action is always one line in `TODAY.md`.

**Where this week sits.**
- ← **From:** v4 and v5 thinking; your August 2026 synthesis ("I build instruments, never products").
- → **Feeds:** every week. Week 1 needs `batchlens` and `TODAY.md`; week 16 needs ORCID; week 20 needs the profile you start today.
- ⟳ **Returns in:** every gate (4, 8, 13, 16, 20...) when you update the headline and the gauge.

**Goals.**
- 🎯 **This week:** the cockpit exists, Monday's `TODAY.md` line is written, post 0 is public.
- 🏁 **Next milestone:** Gate 4, public and reproducible `batchlens v0.1`.
- 🌟 **Final (week 40):** a public trail of 40 weeks that turns "applied AI engineer" into "AI scientist for biotech" in the eyes of people who never met you.

**The plan (v5, unchanged).**
- **Goal:** everything exists and is empty, in the right place.
- **Build:** [ ] `cockpit` repo with the five files; paste Part 6 and 7 into `PROGRESS.md`. [ ] `labloop-notes`, `batchlens`, `fermlit` repos with README skeletons. [ ] Local environment (5.3). [ ] LinkedIn headline and About (Part 9.1). [ ] NotebookLM notebooks created.
- **Tue/Thu:** none. Read Part 0 twice.
- **Sun:** L-Sun post 0: "40 weeks from applied-AI engineer to published AI scientist for biotech, in public, at 10 h/week. First: a fermentation in 80 lines." Write the six `TODAY.md` lines for week 1.
- **Only one thing:** the `cockpit` repo with `TODAY.md` filled for Monday.

**Learn it.**
- ▶ **Core** (no study blocks this week; read these in the spare 30 minutes)
  - *Learning How to Learn* (Coursera, McMaster/UCSD), week 1 lessons on **focused vs diffuse mode** and **chunking** (~40 min). Why: it explains, with neuroscience, why the rituals in Part 0 work.
  - Richard Hamming, *You and Your Research* (1986 talk; transcript free online, video on YouTube). Why: the classic talk on doing important work instead of busy work, from a Bell Labs scientist. Read the section on "important problems" (~20 min).
- ＋ **Deeper**
  - Uri Alon, "How to choose a good scientific problem", *Molecular Cell* 35 (2009) 726-728. Two pages. The "feasibility vs interest" chart you will reuse in week 21.
  - *Good enough practices in scientific computing*, Wilson et al., *PLoS Computational Biology* 2017. The repo habits you are setting up today, justified.
  - *The Turing Way* handbook — book.the-turing-way.org — chapter "Reproducible research", skim.
  - Elliot Hershberg's *Century of Biology* newsletter (search the title) — one essay, for the big picture of where biotech is going.

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| astral-sh/uv | ★ 89.8k | `uv init batchlens && uv add numpy scipy pandas matplotlib`; your environment manager for all 40 weeks |
| ollama/ollama | ★ 180.8k | `ollama run qwen2.5:7b-instruct-q4_K_M "explain Monod growth in two sentences"`; confirm it fits the 3050 |
| drivendataorg/cookiecutter-data-science | ★ 10.1k | Read the folder structure in the README; copy only the ideas (`data/raw` never committed, `reports/figures`) into your skeletons |
| the-turing-way/the-turing-way | ★ 2.2k | Browse the "Guide for reproducible research" table of contents; save two checklists to `PARKING.md` |

**NotebookLM kit · `LABLOOP-W00 · Cockpit and positioning`**
- **Load (7):** (1) this chapter as text; (2) Part 0 of this document as text; (3) Hamming transcript (web); (4) Alon 2009 PDF; (5) Wilson et al. 2017 (PLoS, web); (6) *Learning How to Learn* focused/diffuse lecture (YouTube version, search the title); (7) your LinkedIn About draft.
- **Focus line for the Studio set:** `Focus: how an experienced program manager should structure 40 weeks of learning with ADHD-friendly rituals, and what "important problems" means for a newcomer to biotech.`
- **Custom question:** `Using Hamming and Alon, critique my thesis choice (FermLit) as an "important and feasible problem" for someone with 10 hours a week. Be specific.`

**Mentor check.** *"I know exactly what I will open on Monday and why every file in my cockpit exists."*

---

### Week 1 · A tank in 80 lines

**Act I · Threads: B, M, BL · Next gate: 4**

**Welcome.** Imagine a restaurant kitchen at 7 pm. The chef (the microbe) is cooking. The pantry (the substrate, sugar) is being emptied. The dishes going out (the product) pile up. The air in the kitchen (dissolved oxygen) gets thinner as more cooks join. At first nothing happens (the cooks arrive and put their aprons on: the lag phase), then orders fly (exponential growth), then the pantry runs low and the pace flattens (stationary phase). A fermentation tank is that kitchen, and this week you write its rules as four equations. The Monod equation is simply "cooks work faster when the pantry is full, up to a limit". Once your kitchen runs in Python, you will stream it over MQTT so it looks exactly like a real plant on a Grafana screen. On Tuesday you learn what the cooks are made of (proteins, enzymes). On Thursday you learn the language of the recipe book (vectors and matrices), which every later model is written in.

**Where this week sits.**
- ← **From:** week 0 cockpit and `batchlens` repo.
- → **Feeds:** week 2 (you will recognise X, S, P and DO in 100 real batches), week 3 (your fake fault becomes the first thing your alarms must catch), week 9-11 (the same equations reappear as mass balances in NPTEL), week 12 (your Monod model becomes the "mechanistic half" of the hybrid model), week 17 (Paper 2 learning curves).
- ⟳ **Returns in:** week 12 and Paper 2. The 80 lines you write this week are literally the physics inside Paper 2's hybrid model.

**Goals.**
- 🎯 **This week:** a simulator that shows lag, log and stationary phases, streamed live, plus 20 seeded batches with one labelled fault.
- 🏁 **Next milestone:** Gate 4 (`batchlens v0.1` with DOI).
- 🌟 **Final:** Paper 2 and JOSS: a hybrid soft sensor whose mechanistic core is the model you start today.

**The plan (v5, unchanged).**
- **Goal:** a fed-batch simulator that streams like a plant.
- **Tue (B1):** Biology 2e ch. 3 + 6.5. Notes.
- **Thu (M1):** 3Blue1Brown linear algebra ch. 1-7. Notes. Last 30 min: founder paper 1 (Michroma or Levya author).
- **Build:** [ ] Mon: `batchlens/sim/monod.py`: ODEs for biomass X, substrate S, product P, dissolved oxygen; batch and fed-batch. Done when a plot shows lag, log, stationary. [ ] Wed: MQTT publisher (`paho-mqtt`) sending one sample per simulated minute to the Debian box; InfluxDB + Grafana dashboard. Done when you see the curve live. [ ] Fri: 20 batches with random noise and one injected fault (feed pump stops); saved as CSV. Done when `make sim` writes 20 files.
- **Sun:** L-Sun: "A fermentation in 80 lines, streamed like a plant" with the Grafana screenshot. L-Wed (Wednesday): "what a blind bioreactor costs" (archetype note).
- **Only one thing:** `monod.py` with the plot.

**Learn it.**
- ▶ **Core**
  - Tue · OpenStax *Biology 2e* 6.5 Enzymes — https://openstax.org/books/biology-2e/pages/6-5-enzymes (25 min). Stop when you can explain the Monod term.
  - Thu · 3Blue1Brown *Essence of Linear Algebra*, ch. 1-3 (YouTube playlist, ~30 min).
- ＋ **Deeper**
  - OpenStax *Biology 2e* ch. 3 (macromolecules: carbohydrates, lipids, proteins) — https://openstax.org/books/biology-2e/pages/3-introduction. Moved out of core on 2026-09-29.
  - Crash Course Biology #3 "Biological Molecules - You Are What You Eat" (YouTube, 14 min) and Khan Academy "Enzymes and activation energy" (10 min).
  - 3Blue1Brown *Essence of Linear Algebra*, ch. 4-7 (matrix multiplication, 3D, determinant, inverse). Not needed until later.
  - MIT OCW 18.06 *Linear Algebra* (Gilbert Strang), lectures 1-3 — ocw.mit.edu/courses/18-06-linear-algebra-spring-2010.
  - NPTEL *Bioreactors*, lecture 1 (introduction) — nptel.ac.in/courses/102106053. A preview of week 9, useful if the kitchen analogy needs an engineer's version.
  - LearnChemE (YouTube): search "Monod" and "batch reactor design equation" for 5-minute screencasts.
  - SciPy documentation for `scipy.integrate.solve_ivp` (the solver you will use); read the "Examples" section.
  - Pioreactor project docs (pioreactor.com) — see a real, open-source mini bioreactor with optical density and temperature sensors; it makes "online variables" tangible.

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| jakevdp/PythonDataScienceHandbook | ★ 49.9k | Colab: notebook `02.00-Introduction-to-NumPy` and `02.02-The-Basics-Of-NumPy-Arrays`; re-do the 3Blue1Brown matrix-on-vector examples in NumPy |
| fastai/numerical-linear-algebra | ★ 11.0k | Colab: lesson 1 notebook; skim, run the first cells. Linear algebra as computation, not proofs |
| scipy/scipy | ★ 15.0k | Copy the `solve_ivp` Lotka-Volterra example; replace predator-prey with biomass-substrate. This is 70% of your Monday task |
| eclipse-paho/paho.mqtt.python | ★ 2.4k | Run the "publish single message" example against Mosquitto on the Debian box |
| Pioreactor/pioreactor | ★ 149 (niche, open hardware) | Read `README` and the docs page on growth-rate estimation; no hardware needed. Shows how real online signals look |

**NotebookLM kit · `LABLOOP-W01 · A tank in 80 lines`**
- **Load (10):** (1) this chapter; (2) OpenStax *Biology 2e* ch. 3 web page; (3) ch. 6.5 web page; (4) Crash Course #3 YouTube; (5) Khan enzymes page; (6) 3Blue1Brown linear algebra ch. 1 and ch. 3 YouTube links; (7) SciPy `solve_ivp` docs page; (8) NPTEL Bioreactors lecture 1 PDF or video; (9) your `notes/bio/B1.md`; (10) your `sim/monod.py` pasted as text after Monday.
- **Focus line:** `Focus: the Monod growth model as a restaurant kitchen (biomass, substrate, product, oxygen), what an enzyme is, and why a matrix is a machine that transforms vectors.`
- **Custom question:** `Look at my monod.py. Which biological assumptions does it make silently (lag phase, maintenance, oxygen limitation), and which one would a bioprocess engineer challenge first?`

**Mentor check.** *"I can draw the growth curve, point at each phase, and show the line of code that creates it."*

---

### Week 2 · One hundred real batches

**Act I · Threads: B, M, BL · Next gate: 4**

**Welcome.** Last week you flew the simulator. This week you open the black boxes of 100 real flights. IndPenSim is a benchmark built by researchers from an industrial-scale penicillin fermentation: 100 batches, each hours long, with temperature, pH, oxygen, feeds, and even Raman spectra, some of them with faults. The first feeling will be overwhelm: dozens of variables, thousands of time points. Your job is the job you have done for 25 years with project portfolios: **tidy the data before you judge it.** One long table (batch, time, variable, value). Then overlay ten batches and look. Then PCA, which is like compressing a 30-page report into a two-line executive summary without losing the story: most batches cluster together, and the troubled ones sit visibly apart. On Tuesday you learn why a cell needs sugar and oxygen at all (respiration: the cell's power plant). On Thursday you learn what a gradient is (the slope of the hill a model walks down), which is how every model you will train "learns".

**Where this week sits.**
- ← **From:** week 1 (you know what X, S, P and DO mean because you invented them).
- → **Feeds:** week 3 (PCA scores become Hotelling T² alarms), week 4 (hours-to-detection needs aligned batches), week 8 (soft sensor uses the tidy parquet), weeks 12 and 17 (Paper 2 is built on this dataset).
- ⟳ **Returns in:** Paper 2 "Data" section; you will cite Goldrick et al. 2015 and 2019.

**Goals.**
- 🎯 **This week:** IndPenSim as a tidy parquet that loads in under a minute; an overlay plot; faulty batches visible in PCA.
- 🏁 **Next milestone:** Gate 4.
- 🌟 **Final:** the dataset under Paper 2; your first Medium thesis piece (M1) positions you before any paper exists.

**The plan (v5, unchanged).**
- **Goal:** IndPenSim loaded, understood, and one golden-batch picture.
- **Tue (B2):** Biology 2e ch. 4, 5.1-5.2, 7.
- **Thu (M2):** 3Blue1Brown calculus ch. 1-4 + neural networks ch. 1-4. Founder paper 2.
- **Build:** [ ] Mon: `data/download.py` for IndPenSim with checksums; load into a tidy long table (batch, time, variable, value). Done when 100 batches load in under a minute from parquet. [ ] Wed: batch unfolding and time alignment; plot every variable for 10 batches overlaid. Done when you can name the 5 variables that matter. [ ] Fri: PCA on the "normal" batches, scores plot, first look at the faulty ones. Done when the faulty batches are visibly outside.
- **Sun:** L-Sun with the overlay plot. **M1** (Medium, thesis piece): "Biotech startups don't have an AI problem" (reuse v4 section 1, archetypes with counts, no names).
- **Only one thing:** the tidy parquet of IndPenSim.

**Learn it.**
- ▶ **Core**
  - Tue · Crash Course Biology #7 "ATP & Respiration" (YouTube, 13 min). Stop when you can say why DO falls as biomass grows.
  - Thu · 3Blue1Brown *Essence of Calculus* ch. 1-2 (~25 min).
  - Build-block reading · Goldrick et al. 2015, "The development of an industrial-scale fed-batch fermentation simulation", *J. Biotechnology* 193:70-82, abstract, figures and section 2; Goldrick et al. 2019, "Modern day monitoring and control challenges outlined on an industrial-scale benchmark fermentation process", *Computers & Chemical Engineering* 130, abstract and the dataset section.
  - StatQuest "Principal Component Analysis (PCA), Step-by-Step" (YouTube, 22 min) before Friday's PCA.
- ＋ **Deeper**
  - OpenStax *Biology 2e* ch. 7 (cellular respiration), ch. 4 and 5.1-5.2. Moved out of core on 2026-09-29.
  - 3Blue1Brown *Essence of Calculus* ch. 3-4 and *Neural Networks* ch. 1-2 (Karpathy covers networks in week 8).
  - Steve Brunton, "Singular Value Decomposition (SVD)" series, first 3 videos (YouTube). PCA's engine, told by an engineer.
  - Ninja Nerd, glycolysis and Krebs cycle lectures (YouTube), if respiration still feels abstract.
  - *Mathematics for Machine Learning* (Deisenroth, Faisal, Ong), ch. 10 "Dimensionality Reduction with PCA" — mml-book.github.io (free PDF).
  - industrialpenicillinsimulation.com — the simulator site (verify link and license before download).

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| StephenGoldie/indpensim-notebook | ★ 29 (niche; appears to be from the IndPenSim team — verify author) | Open `IndPenSim_Summary_data_notebook_V3.ipynb` in Colab; compare their variable list with your tidy table |
| jakevdp/PythonDataScienceHandbook | ★ 49.9k | Colab: `03.08-Aggregation-and-Grouping` and `05.09-Principal-Component-Analysis` |
| scikit-learn/scikit-learn | ★ 67.2k | Run the gallery example "PCA example with Iris Data-set", then swap Iris for 20 IndPenSim batches unfolded |
| pandas-dev/pandas / pola-rs/polars | ★ 49.7k / ★ 39.7k | `melt` your wide CSVs to long format, save as parquet; time both libraries on the 100 batches |

**NotebookLM kit · `LABLOOP-W02 · One hundred real batches`**
- **Load (10):** (1) this chapter; (2) Goldrick 2015 PDF; (3) Goldrick 2019 PDF; (4) OpenStax ch. 7 web page; (5) Crash Course #7 YouTube; (6) 3Blue1Brown calculus ch. 1 and ch. 2; (7) StatQuest PCA YouTube; (8) IndPenSim notebook README (GitHub page); (9) your overlay plot + the list of "5 variables that matter" as text; (10) `notes/ml/M2.md`.
- **Focus line:** `Focus: what IndPenSim measures and why, how cellular respiration explains the oxygen and substrate curves, and PCA as an executive summary of 100 batches.`
- **Studio extra (Data Table):** `Table of every IndPenSim variable: name, unit, online or offline, what biology or engineering it reflects.`
- **Custom question:** `From the Goldrick papers, which faults exist in the 100 batches and how are they labelled? Quote the sentences.`

**Mentor check.** *"I can explain what a faulty batch looks like in a PCA scores plot and why respiration makes oxygen fall when biomass rises."*

---

### Week 3 · Golden batch, with statistics

**Act I · Threads: B, M, BL · Next gate: 4**

**Welcome.** Every plant manager has a "golden batch": the run that went perfectly and that everyone secretly compares against. The problem is that a single golden batch is an anecdote. This week you replace the anecdote with an envelope: what 80 normal batches look like together, with statistical limits instead of someone's gut thresholds. The analogy from your world is the portfolio dashboard. A single project running 5% late is noise; a project that is late *and* over budget *and* losing staff at the same time is a pattern. Hotelling T² is exactly that: it watches many variables together and raises its hand when their *combination* is unusual, even if each alone looks fine. SPE is the other hand: it notices when a batch does something the normal batches never did at all. On Tuesday, you study microbial growth properly and discover that your Monod model already explains the log phase. On Thursday, you start *An Introduction to Statistical Learning*, the most-used ML textbook in the world, written by Stanford statisticians.

**Where this week sits.**
- ← **From:** week 2 PCA on normal batches; week 1 injected fault.
- → **Feeds:** week 4 (hours-to-detection and false-alarm rate), week 8 (technical report "golden batch" section), week 17 (hours-to-detection per model in Paper 2), week 33-35 (batchlens alarms trigger labdocs deviation drafts).
- ⟳ **Returns in:** every Loop Audit (week 29): "how early would you have known?"

**Goals.**
- 🎯 **This week:** T² and SPE limits computed from data; `batchlens monitor --batch 87` alarms on a faulty batch and not on a normal one; CI green.
- 🏁 **Next milestone:** Gate 4.
- 🌟 **Final:** the early-warning headline number that opens founder conversations in archetype A.

**The plan (v5, unchanged).**
- **Goal:** control limits that are statistics, not thresholds.
- **Tue (B3):** Microbiology ch. 9. Draw the growth curve.
- **Thu (M3):** ISLP ch. 2-3. Redo Lab 3 on IndPenSim.
- **Build:** [ ] Mon: Hotelling T² and SPE with limits from the normal batches (NIST ch. 6.5 as reference). Done when limits are computed, not typed. [ ] Wed: `batchlens monitor --batch 87` replays a batch and prints alarms with time stamps. Done when a faulty batch alarms and a normal one does not. [ ] Fri: tests for the two commands; `make report FAKE=1` works on simulator data. Done when CI is green.
- **Sun:** L-Sun: "Golden batch: what 80 normal fermentations look like as one envelope". L-Wed.
- **Only one thing:** T² limits computed.

**Learn it.**
- ▶ **Core**
  - Tue · OpenStax *Microbiology* ch. 9 "Microbial Growth", growth-curve section only — openstax.org/details/books/microbiology (30 min); draw the curve by hand.
  - Thu · 3Blue1Brown "Eigenvectors and eigenvalues" (*Essence of Linear Algebra* ch. 14, YouTube, 17 min): what PCA actually found last Friday.
  - Build-block reading · NIST/SEMATECH e-Handbook ch. 6.5 "Tutorials: multivariate control charts" — https://www.itl.nist.gov/div898/handbook/pmc/section3/pmc34.htm (sections 6.3.4 and 6.5.4.3).
  - StatQuest "Linear Regression, Clearly Explained" and "Machine Learning Fundamentals: Bias and Variance" (YouTube, 25 min total).
- ＋ **Deeper**
  - ISLP ch. 2 and 3.1-3.3 — statlearning.com (free PDF) + the matching Stanford *Statistical Learning with Python* videos. Moved out of core on 2026-09-29.
  - Nomikos & MacGregor (1995), "Multivariate SPC charts for monitoring batch processes", *Technometrics* 37(1):41-59 — the founding paper of batch MSPC. Read abstract + figures.
  - Kourti & MacGregor (1995), "Process analysis, monitoring and diagnosis, using multivariate projection methods", *Chemometrics and Intelligent Laboratory Systems* 28:3-21 (verify).
  - Coursera *Introduction to Statistics* (Stanford, Walther), module on regression, if ISLP ch. 3 feels steep.
  - Crash Course Statistics (YouTube), episodes on the normal distribution and control charts' ancestor, the z-score.

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| intro-stat-learning/ISLP_labs | ★ 1.4k (official, Stanford authors) | Colab: `Ch02-statlearn-lab.ipynb`, then `Ch03-linreg-lab.ipynb`; redo the regression on two IndPenSim variables |
| erdogant/pca | ★ 354 | Its README shows Hotelling T² and SPE outlier detection on top of PCA; run the example, then compare its limits with yours (never ship its numbers without understanding them) |
| salvadorgarciamunoz/pyphi | ★ 61 (niche; by Sal García-Muñoz, a pharma MSPC practitioner) | Read the README and one example of PCA/PLS with T² and SPE; a pharma-industry reference implementation to sanity-check yours |
| statsmodels/statsmodels | ★ 11.6k | Fit an OLS on two variables and read the summary table line by line with ISLP ch. 3 open |

**NotebookLM kit · `LABLOOP-W03 · Golden batch, with statistics`**
- **Load (10):** (1) this chapter; (2) OpenStax *Microbiology* ch. 9 web; (3) ISLP ch. 2-3 PDF (or the chapter pages); (4) Stanford Statistical Learning ch. 2 and 3 videos (YouTube); (5) NIST handbook 6.5 web pages; (6) StatQuest regression + bias-variance (YouTube); (7) Nomikos & MacGregor 1995 abstract (web); (8) `erdogant/pca` GitHub page; (9) your T² plot description and limits as text; (10) `QUESTIONS.md`.
- **Focus line:** `Focus: from a golden batch anecdote to a statistical envelope; Hotelling T² and SPE as two different ways a batch can be abnormal; the bias-variance trade-off.`
- **Custom question:** `Explain the difference between T² and SPE alarms using a portfolio-dashboard analogy, then tell me which of the two should fire first for a feed-pump failure and why.`

**Mentor check.** *"I can explain to a plant manager why their golden batch is not a control limit, and what T² and SPE each catch."*

---

### Week 4 · L1: public and reproducible

**Act I (finale) · Threads: B, M, BL, R · GATE 4**

**Welcome.** In your corporate life, a deliverable is done when a steering committee signs it. In science, a result is done when a stranger can regenerate it without calling you. This week is your first licence exam. You will turn your alarms into two honest numbers (how many hours before harvest you detect a fault; how often you cry wolf on a normal batch), put them in a README with one command that regenerates them, and mint a DOI through Zenodo so the release becomes citable forever. There is a trap here that catches most people moving from engineering into science: **leakage across batches.** If time points from the same batch sneak into both training and testing, your model "remembers" the batch instead of learning the process, like a candidate who saw the exam questions. On Thursday you learn cross-validation done right, and you implement batch-wise CV so that trap is closed before Paper 2 depends on it. On Tuesday, microbial metabolism: aerobic vs anaerobic, the reason some tanks bubble air and others do not.

**Where this week sits.**
- ← **From:** weeks 1-3 (simulator, tidy data, alarms).
- → **Feeds:** week 8 (technical report cites v0.1), week 12 and 17 (batch-wise CV is the backbone of the learning curves), week 36 (JOSS needs a clean release history).
- ⟳ **Returns in:** every release: v0.2 (wk 8), v0.3 (wk 17).

**Goals.**
- 🎯 **This week:** `batchlens v0.1` released with a DOI, headline number in README, `make report` reproduces it, batch-wise CV implemented.
- 🏁 **This gate (4):** L1 public and reproducible + L2 citable (DOI).
- 🌟 **Final:** the release history JOSS reviewers will read in week 36.

**The plan (v5, unchanged).**
- **Goal:** first public number and first DOI.
- **Tue (B4):** Microbiology ch. 8.1-8.4.
- **Thu (M4):** ISLP ch. 5. Implement batch-wise CV.
- **Build:** [ ] Mon: hours-to-detection on every faulty batch; false-alarm rate on normal batches; one table. Done when the table regenerates with `make report`. [ ] Wed: README with the headline number, `CITATION.cff`, Zenodo integration. Done when the DOI badge renders. [ ] Fri: release `batchlens v0.1`. Send the repo to your Sumble friend for a 30-minute code review.
- **Sun:** L-Sun: "Deviation detected X hours before harvest on N faulty batches, Y% false alarms". **Z1**. Gauge: **L1**, and L2 for the DOI.
- **Only one thing:** the release with the DOI.

**Learn it.**
- ▶ **Core**
  - Tue · Khan Academy "Fermentation and anaerobic respiration" (10 min).
  - Thu · StatQuest "Machine Learning Fundamentals: Cross Validation" (YouTube, 6 min); Kapoor & Narayanan (2023), "Leakage and the reproducibility crisis in machine-learning-based science", *Patterns* 4(9) — the leakage taxonomy only (20 min).
  - Build-block reading · GitHub Docs "Referencing and citing content" (the Zenodo integration page) and the `CITATION.cff` format page (citation-file-format.github.io).
- ＋ **Deeper**
  - OpenStax *Microbiology* ch. 8.1-8.4 "Microbial Metabolism"; ISLP ch. 5 with the Stanford ch. 5 videos; StatQuest "Bootstrapping Main Ideas!!!". Moved out of core on 2026-09-29.
  - Sandve et al. (2013), "Ten simple rules for reproducible computational research", *PLoS Computational Biology* 9(10):e1003285.
  - *The Turing Way*, chapter on "Version control" and "Research compendia".

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| intro-stat-learning/ISLP_labs | ★ 1.4k | Colab: `Ch05-resample-lab.ipynb`; then rewrite one cell using `GroupKFold` with batch as the group |
| scikit-learn/scikit-learn | ★ 67.2k | User guide "Cross-validation: evaluating estimator performance", section on group-wise splitters; run the visualization example "Visualizing cross-validation behavior" |
| google-research/tuning_playbook | ★ 30.3k | Read only the section on evaluation and "incremental tuning"; one page of habits you will reuse in weeks 8 and 17 |
| drivendataorg/cookiecutter-data-science | ★ 10.1k | Compare your `Makefile` with its conventions; add `make report FAKE=1` |

**NotebookLM kit · `LABLOOP-W04 · L1 public and reproducible`** (gate week: add Video Overview + Slide Deck)
- **Load (10):** (1) this chapter; (2) ISLP ch. 5 PDF; (3) Stanford ch. 5 videos; (4) StatQuest CV + bootstrap; (5) Kapoor & Narayanan 2023 (web); (6) Sandve et al. 2013 (web); (7) OpenStax *Microbiology* ch. 8 web; (8) GitHub/Zenodo citing guide page; (9) your README draft with the headline number; (10) `LOG.md` weeks 1-4.
- **Focus line:** `Focus: why reproducibility and batch-wise cross-validation are what turn a demo into a result; aerobic vs anaerobic metabolism in a tank.`
- **Slide Deck prompt:** `Five slides for a founder: the problem (deviations found at harvest), the data (IndPenSim), the method (golden envelope + T²/SPE), the number (hours to detection, false alarms), what it does not do.`
- **Custom question:** `Audit my README like a JOSS reviewer: what would you need to reproduce the headline number, and what is missing?`

**Mentor check.** *"A stranger can clone batchlens, run one command, and get my number. I can explain why batch-wise CV is not optional."*

---

# ACT II · The reading robot (weeks 5-8)

> **Act II story.** Picture a consulting firm that hires a brilliant, impossibly fast junior analyst. On day one they could read 2,000 client reports over a weekend. No partner would send those summaries to a client without first grading the analyst's work on 50 reports the partner read personally. That is Act II. FermLit is the analyst (an LLM agent). The gold set is the partner's answer key. The F1 table is the performance review. And the technical report at week 8 is the first document you sign with your name as a scientist.

---

### Week 5 · FermLit begins: the corpus

**Act II · Threads: B, M, FL · Next gate: 8**

**Welcome.** Every serious research project starts with an uncomfortable question: *has somebody already done this?* In corporate strategy you would call it a competitive scan. In science it is the related-work search, and skipping it is the most common way a newcomer wastes six months. So the first thing FermLit does is look for its own twin. If a fermentation-literature dataset already exists, wonderful: you extend it and cite it, which is still a paper. Then you go fishing. Europe PMC is an ocean of open-access papers; your query is the net; the 500 XML files are the catch. And then the schema, which is the most "program manager" thing a scientist ever writes: a contract that says exactly which fields every paper must yield, in which unit, with which sentence as evidence. On Tuesday, the central dogma: DNA → RNA → protein, so that "the strain overexpresses gene X" stops being noise. On Thursday, regularization and trees: why boosted trees win on tabular data, which you will need for the soft sensor.

**Where this week sits.**
- ← **From:** week 4 reproducibility habits (manifest, checksums); your v3/Vara experience with LLM pipelines.
- → **Feeds:** week 6 (the schema becomes annotation guidelines), week 7 (the agent fills the schema), week 9 (corpus scales to 2,000), week 14 (`RELATED.md` becomes Paper 1's related work).
- ⟳ **Returns in:** Paper 1 Methods: "Corpus construction" and "Schema".

**Goals.**
- 🎯 **This week:** `RELATED.md` with ≥ 10 verified references; 500 open-access XMLs + manifest; `fermlit.schema.json` v0 that fits three hand-annotated papers.
- 🏁 **Next milestone:** Gate 8 (technical report on Zenodo, F1 table public).
- 🌟 **Final:** Paper 1 on bioRxiv (wk 16) and a dataset other scientists cite (L6).

**The plan (v5, unchanged).**
- **Goal:** 500 open-access fermentation papers on disk and a written schema.
- **Tue (B5):** Biology 2e ch. 14-15; Rosalind 1-4.
- **Thu (M5):** ISLP ch. 6.1-6.2, 8. LightGBM soft sensor v0.
- **Build:** [ ] Mon: Prompt P8 with Semantic Scholar: does a fermentation literature dataset already exist? Write `fermlit/RELATED.md` with what you found (this decides "novel" vs "extension"). Done when the file has ≥ 10 verified references. [ ] Wed: Europe PMC query (`"fed-batch" OR "fermentation" AND ("titer" OR "titre") AND OPEN_ACCESS:y`, refine), download full-text XML for 500 papers. Done when 500 XML files and a manifest exist. [ ] Fri: `schema/fermlit.schema.json` v0: organism, strain, product, reactor type, working volume, mode (batch/fed-batch/continuous), medium carbon source, temperature, pH, DO setpoint, feed strategy, duration, final titer (g/L), yield, productivity, scale, plus a provenance sentence per field. Done when three papers annotated by hand against it fit.
- **Sun:** L-Sun: "500 fermentation papers, one schema: what I want the robot to read". L-Wed.
- **Only one thing:** `RELATED.md`. If someone did it already, you must know now.

**Learn it.**
- ▶ **Core**
  - Tue · OpenStax *Biology 2e* ch. 14 (DNA structure and function) + ch. 15 (genes and proteins), sections 14.1-14.3, 15.1-15.5 (60 min).
  - Tue · Crash Course Biology #11 "DNA, Hot Pockets, & The Longest Word Ever" (YouTube, 13 min); Rosalind problems 1-4 (rosalind.info).
  - Thu · ISLP ch. 6.1-6.2 (subset selection, ridge, lasso) and ch. 8 (trees, bagging, boosting) (75 min, skim the proofs).
  - StatQuest "Regularization Part 1: Ridge (L2) Regression" and "Gradient Boost Part 1: Regression Main Ideas" (YouTube, 25 min).
  - Build-block reading · Europe PMC REST API docs — europepmc.org/RestfulWebService (search syntax, `OPEN_ACCESS:y`, `fullTextXML`).
- ＋ **Deeper**
  - Dagdelen et al. (2024), "Structured information extraction from scientific text with large language models", *Nature Communications* 15:1418. The closest methodological cousin of FermLit (materials science). Read it this week; you will cite it.
  - Polak & Morgan (2024), "Extracting accurate materials data from research papers with conversational language models and prompt engineering", *Nature Communications* 15:1569.
  - Uri Alon (2009), "How to choose a good scientific problem" (re-read from week 0, now with FermLit in mind).
  - Coursera *Biology Meets Programming: Bioinformatics for Beginners* (UCSD), week 1, if Rosalind hooks you.
  - MIT 7.00x (edX), molecular biology unit, for Lander's storytelling of the central dogma.

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| J535D165/pyalex | ★ 408 | Query OpenAlex for works on "fed-batch fermentation titer" per year; plot counts. Tells you the size of the ocean before you fish |
| danielnsilva/semanticscholar | ★ 480 | Search Semantic Scholar for "literature-derived fermentation dataset"; save the top 20 to `RELATED.md` with citation counts |
| titipata/pubmed_parser | ★ 738 | Parse one PMC OA XML into title, abstract, methods paragraphs; this becomes `corpus/parse.py` |
| google/langextract | ★ 38.6k | Read the README and run the quickstart example: extraction with **source grounding** (every value mapped to its exact span). This is the provenance idea FermLit needs, already implemented |
| intro-stat-learning/ISLP_labs | ★ 1.4k | Colab: `Ch08-baggboost-lab.ipynb` (trees and boosting) |

**NotebookLM kit · `LABLOOP-W05 · FermLit corpus and schema`**
- **Load (12):** (1) this chapter; (2) Dagdelen 2024 (Nat Commun, web/PDF); (3) Polak & Morgan 2024 (web/PDF); (4) Alon 2009 PDF; (5) Europe PMC REST API page; (6) OpenStax ch. 15 web; (7) Crash Course #11; (8) ISLP ch. 8 PDF; (9) StatQuest gradient boost; (10) LangExtract GitHub page; (11) your `RELATED.md`; (12) `fermlit.schema.json` pasted as text.
- **Focus line:** `Focus: how LLMs extract structured data from scientific papers with provenance, what a good extraction schema looks like, the central dogma, and why boosted trees work on tabular data.`
- **Studio extra (Data Table):** `From Dagdelen 2024 and Polak & Morgan 2024: task, domain, number of papers, model, evaluation metric, reported accuracy, how they built the gold set.`
- **Custom question:** `Compare my schema with the extraction targets in Dagdelen 2024 and Polak & Morgan 2024. Which of my fields will be hardest to extract reliably and why?`

**Mentor check.** *"I know whether FermLit is novel or an extension, and I can defend every field in my schema."*

---

### Week 6 · The gold set (the most important week of Half 1)

**Act II · Threads: B, M, FL · Next gate: 8**

**Welcome.** Here is a story from auditing. A company hires a new automated invoice-checking system. The vendor claims 98% accuracy. The CFO asks one question: "98% against what?" If nobody checked invoices by hand first, the number means nothing. This week you become that CFO. You read 50 fermentation papers yourself and write down, for each field, the value and the exact sentence that proves it. It is slow, it is humbling, and it is the single thing that makes FermLit science instead of a demo. Two things will happen: you will discover that papers are ambiguous in ways no schema anticipated (a range of temperatures, three strains in one table, titer in "mg per litre of broth"), and you will write those decisions into guidelines, which is what makes a second person able to repeat your work. That second person is the other job this week: recruit an independent annotator so week 10 can measure agreement. On Tuesday, genetic engineering: what "recombinant *E. coli*" means. On Thursday, probability: the language in which every uncertainty in your papers will be written.

**Where this week sits.**
- ← **From:** week 5 schema and corpus.
- → **Feeds:** week 7 (the agent is scored against these 50), week 10 (kappa with the second annotator), week 14 (Paper 1 Methods: "Gold standard"), week 34 (labdocs uses the same method with 60 Q&A).
- ⟳ **Returns in:** every evaluation you will ever sell. The gold-set habit is your professional signature.

**Goals.**
- 🎯 **This week:** 50 gold JSONs with provenance sentences; `GUIDELINES.md`; one second annotator said yes.
- 🏁 **Next milestone:** Gate 8.
- 🌟 **Final:** the human-validation claim in Paper 1's title, and the calibrated-evaluation reputation that sells Loop Audits.

**The plan (v5, unchanged).**
- **Goal:** 50 papers annotated by hand, guidelines written, second annotator recruited.
- **Tue (B6):** Microbiology ch. 12; Rosalind 5-8.
- **Thu (M6):** Murphy ch. 2 or Statistical Rethinking L1-2.
- **Build:** [ ] Mon: `annotate/GUIDELINES.md` with Prompt P7. Annotate 15 papers. Done when 15 gold JSONs with provenance sentences exist. [ ] Wed: annotate 20 more; refine guidelines with every ambiguity you hit. [ ] Fri: annotate 15 more (50 total). Message two people to annotate 30 of them independently (a founder-scientist, a biology student, or a paid postdoc). Done when one has said yes.
- **Sun:** L-Sun: "I read 50 fermentation papers by hand this week. Here is what they never tell you" (first hint of the reporting finding). L-Wed.
- **Only one thing:** 50 gold JSONs.

**Learn it.**
- ▶ **Core**
  - Tue · OpenStax *Microbiology* ch. 12.1-12.2 "Modern Applications of Microbial Genetics" (60 min); Rosalind problems 5-8.
  - Thu · Statistical Rethinking 2024, lectures 1-2 (YouTube, from github.com/rmcelreath/stat_rethinking_2024) **or** Murphy, *Probabilistic Machine Learning: An Introduction*, ch. 2 — probml.github.io/pml-book/book1.html (75 min). Pick Rethinking if you want stories; Murphy if you want equations.
  - Build-block reading · Datasheets for Datasets: Gebru et al. (2021), *Communications of the ACM* 64(12):86-92. The questions your guidelines and dataset card must answer.
- ＋ **Deeper**
  - Artstein & Poesio (2008), "Inter-coder agreement for computational linguistics", *Computational Linguistics* 34(4):555-596, sections 1-2. The bible of annotation agreement.
  - McHugh (2012), "Interrater reliability: the kappa statistic", *Biochemia Medica* 22(3):276-282. Short, open, written for biomedical readers.
  - 3Blue1Brown "Bayes theorem, the geometry of changing beliefs" (YouTube, 15 min).
  - STRENDA guidelines (Standards for Reporting Enzymology Data) — search "STRENDA guidelines" (verify URL). A real "minimum information" standard; it shows what a reporting checklist looks like and will inspire your completeness analysis.

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| HumanSignal/label-studio | ★ 28.3k | `pip install label-studio`, create a text-span project with 5 FermLit fields; annotate 2 papers to feel the difference from JSON-by-hand. Use it only if it speeds you up |
| doccano/doccano | ★ 10.8k | Lighter alternative for span annotation; compare in 15 minutes, pick one, never revisit |
| argilla-io/argilla | ★ 5.1k | Read the "human feedback on LLM outputs" quickstart; you will use this pattern in week 7 to correct agent outputs |
| pymc-devs/pymc-resources | ★ 2.1k | Colab: Statistical Rethinking chapter 2 notebook (grid approximation); do the globe-tossing update by hand, then in code |

**NotebookLM kit · `LABLOOP-W06 · The human gold standard`**
- **Load (11):** (1) this chapter; (2) Gebru et al. Datasheets (arXiv 1803.09010 PDF); (3) Artstein & Poesio 2008 PDF; (4) McHugh 2012 (web); (5) Rethinking lectures 1-2 YouTube; (6) OpenStax *Microbiology* ch. 12 web; (7) STRENDA guidelines page; (8) your `GUIDELINES.md`; (9) 5 of your gold JSONs pasted; (10) `QUESTIONS.md`; (11) the P7 prompt output.
- **Focus line:** `Focus: why a human gold standard with provenance is the foundation of any trustworthy LLM extraction, how annotation guidelines resolve ambiguity, and probability as the language of uncertainty.`
- **Custom question:** `Read my guidelines and 5 gold JSONs. Find three places where two careful annotators would disagree, and propose a rule for each.`

**Mentor check.** *"I can explain why my 50 hand-annotated papers are worth more than 2,000 machine-extracted ones, and what a second annotator adds."*

---

### Week 7 · The extraction agent, measured

**Act II · Threads: B, M, FL · Next gate: 8**

**Welcome.** Now the junior analyst meets the answer key. You build `extract/agent.py` with two brains, a frontier model through the API and a local Qwen on your 3050, plus a fake brain for tests. Then you do the thing that distinguishes an AI scientist from an AI enthusiast: you **measure it per field**, not as one flattering average. It will be very good at "organism" and bad at "working volume", and that unevenness is the finding. Think of a football team review: a coach does not say "the team is 70% good"; they say "the goalkeeper is excellent and the left back loses every duel". Your error analysis is the match video. On Tuesday you visit the machinery itself: protein structure and enzyme kinetics, and you fit Michaelis-Menten in Python, which is Monod's twin (same curve, one level down: enzyme instead of cell). On Thursday, Bayesian models: how to put error bars on a titer that come from reasoning, not from a formula you copied.

**Where this week sits.**
- ← **From:** week 5 schema, week 6 gold set, your Vara telemetry habits.
- → **Feeds:** week 8 (F1 table public), week 9 (you pick the cheapest backend that meets the F1 floor), week 14 (the calibrated judge for fuzzy fields), week 15 (M15 evaluation writing), week 34-35 (local vs cloud comparison for labdocs).
- ⟳ **Returns in:** Paper 1 Table 2 (per-field precision/recall/F1) and in every Loop Audit prototype.

**Goals.**
- 🎯 **This week:** agent extracts 5 papers with provenance; `eval/score.py` prints per-field P/R/F1 for 50 papers; one round of error analysis logged with before/after.
- 🏁 **Next milestone:** Gate 8.
- 🌟 **Final:** the method section reviewers of Paper 1 will test hardest.

**The plan (v5, unchanged).**
- **Goal:** first per-field precision and recall.
- **Tue (B7):** MIT 7.016 protein lectures; Michaelis-Menten fitted.
- **Thu (M7):** Statistical Rethinking L3-4.
- **Build:** [ ] Mon: `extract/agent.py`: one paper in, schema JSON out with provenance; Claude via API and local Qwen 7B as two backends; `--fake` backend for tests. Done when 5 papers extract. [ ] Wed: `eval/score.py`: per-field exact and fuzzy match against the gold set, precision/recall/F1, confusion by field. Done when the table prints for 50 papers. [ ] Fri: error analysis: the 5 worst fields, why; one round of prompt and schema fixes; re-score. Log both numbers.
- **Sun:** L-Sun: "An LLM read 50 papers; here is where it was wrong" with the per-field table. L-Wed.
- **Only one thing:** the F1 table, whatever it says.

**Learn it.**
- ▶ **Core**
  - Tue · MIT OCW 7.016 lectures on protein structure and enzymes (ocw.mit.edu/courses/7-016-introductory-biology-fall-2018), ~50 min; Khan Academy "Steady states and the Michaelis Menten equation" (MCAT section) (15 min).
  - Thu · Statistical Rethinking 2024 lectures 3-4 (YouTube, 75 min).
  - Build-block reading · Zheng et al. (2023), "Judging LLM-as-a-judge with MT-Bench and Chatbot Arena", NeurIPS Datasets & Benchmarks, sections 1-3 (judge biases you must avoid in week 14).
- ＋ **Deeper**
  - Anthropic Claude Cookbooks (github.com/anthropics/claude-cookbooks) — "extracting structured JSON" and "tool use" recipes.
  - DSPy (Stanford NLP) documentation "Optimizers" — how to tune prompts against a metric instead of by feel.
  - Karpathy, "Deep Dive into LLMs like ChatGPT" (YouTube, 3.5 h; watch the first hour on hallucination and tool use).
  - Ninja Nerd "Enzyme kinetics" lecture (YouTube) for Michaelis-Menten in depth.

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| 567-labs/instructor | ★ 13.9k | Define your schema as a Pydantic model and let instructor enforce valid JSON from Claude and Ollama; kills 80% of parsing bugs |
| google/langextract | ★ 38.6k | Run an example in `examples/notebooks` with a local Ollama model; compare its grounded spans with your provenance sentences |
| stanfordnlp/dspy | ★ 38.0k | Tutorial on "information extraction" in the docs: define a signature, a metric (your F1), and let an optimizer improve the prompt. Only after your manual baseline exists |
| anthropics/claude-cookbooks | ★ 52.7k | Notebook on structured extraction / tool use; borrow the JSON-mode pattern |
| lmfit/lmfit-py | ★ 1.2k | Fit Michaelis-Menten to a small public dataset with confidence intervals on Km and Vmax |

**NotebookLM kit · `LABLOOP-W07 · The extraction agent, measured`**
- **Load (11):** (1) this chapter; (2) Zheng et al. 2023 PDF; (3) Dagdelen 2024 (carry over); (4) MIT 7.016 enzyme lecture video; (5) Khan Michaelis-Menten page; (6) Rethinking lectures 3-4; (7) instructor GitHub page; (8) LangExtract GitHub page; (9) your F1 table as text; (10) your 5-worst-fields error analysis; (11) `LOG.md` of the week.
- **Focus line:** `Focus: measuring LLM extraction per field against a gold set, reading precision and recall, the anatomy of extraction errors, and Michaelis-Menten as Monod's enzyme-level twin.`
- **Custom question:** `Given my per-field F1 table and error notes, which three changes (schema, prompt, or evaluation rule) would most improve trustworthiness without overfitting to the 50 gold papers?`

**Mentor check.** *"I can read my own F1 table aloud, say why the worst field fails, and explain why tuning on the gold set is dangerous."*

---

### Week 8 · GATE 8: technical report and L2

**Act II (finale) · Threads: M, BL, FL, R · GATE 8**

**Welcome.** Eight weeks ago you packed a cockpit. Today you publish your first document with scientific shape. Think of the difference between a slide deck and an audited technical report: the slide deck persuades; the report lets the reader check. Your technical report tells the story of batchlens on IndPenSim (golden batch, deviation detection, and a first soft sensor that predicts titer mid-batch from online variables) and includes a section most consultants never write: *where it fails*. That section is what makes scientists trust the rest. A soft sensor, by the way, is a "virtual instrument": like estimating a car's fuel consumption from speed and RPM when there is no fuel meter. You also send the report to five founders, not to sell, but to ask whether it resembles their tank. On Tuesday you review biology with a quiz; on Thursday you build a neural network's learning engine (backpropagation) by hand with Karpathy, so the word "gradient" becomes muscle memory.

**Where this week sits.**
- ← **From:** weeks 1-4 (batchlens), weeks 5-7 (FermLit F1 table).
- → **Feeds:** week 12 (hybrid model beats or loses to this week's LightGBM/PLS baseline), week 16 (the 5 founders receive the preprint), week 17 (Paper 2 extends this soft sensor), week 12 endorsement asks (the report is your proof of seriousness).
- ⟳ **Returns in:** Paper 2 baselines; the founder conversations that lead to the co-author in week 24.

**Goals.**
- 🎯 **This week:** soft sensor with error at 25/50/75% of batch time; technical report on Zenodo with DOI; 5 founders messaged; F1 table public.
- 🏁 **This gate (8):** L2 citable, solidly.
- 🌟 **Final:** the first link in the chain founder → collaborator → co-author.

**The plan (v5, unchanged).**
- **Goal:** the first artifact with scientific shape.
- **Tue (B8):** review + 2-page summary; Prompt P3 quiz.
- **Thu (M8):** Karpathy micrograd; rewrite from memory Saturday.
- **Build:** [ ] Mon: soft sensor (titer from online variables), batch-wise CV, error at 25/50/75% of batch time; LightGBM vs PLS. Done when the curve regenerates. [ ] Wed: **Technical report 1** (8-10 pages, Quarto): batchlens on IndPenSim: golden batch, deviation detection, soft sensor; methods reproducible; "where it fails". Prompt P9 for the skeleton, P10 for review. [ ] Fri: release `batchlens v0.2` + report PDF on Zenodo. Send the report to 5 archetype-A founders with the Track F outreach line: "would you tell me in 30 minutes whether this resembles your tank?"
- **Sun:** **M2** deep dive: "What 100 open fermentation batches taught me about early deviation detection". **Z2**. L-Sun. Gauge check: **L2**.
- **Gate 8 passes when:** report on Zenodo with DOI, 5 founders messaged, F1 table public.
- **Only one thing:** the report on Zenodo.

**Learn it.**
- ▶ **Core**
  - Tue · Review B1-B7 notes; P3 quiz with the "PI in a lab meeting" persona (60 min).
  - Thu · Karpathy, "The spelled-out intro to neural networks and backpropagation: building micrograd" (YouTube, 2 h 25 min; watch the first 75 min Thursday, the rest Saturday while rewriting from memory).
  - Build-block reading · Luttmann et al. (2012), "Soft sensors in bioprocessing: a status report and recommendations", *Biotechnology Journal* 7(8):1040-1048. The reference framing for your soft-sensor section.
- ＋ **Deeper**
  - Kadlec, Gabrys & Strandt (2009), "Data-driven soft sensors in the process industry", *Computers & Chemical Engineering* 33(4):795-814.
  - 3Blue1Brown *Neural Networks* ch. 3-4 ("Backpropagation, intuitively" and "Backpropagation calculus").
  - Stanford CS231n notes "Backpropagation, intuitions" (cs231n.github.io/optimization-2) — a classic written explanation.
  - Simon Peyton Jones, "How to write a great research paper" (YouTube) — watch now, not in week 18, because the report is a mini-paper.
  - Quarto guide for scientific manuscripts (quarto.org/docs/manuscripts).

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| karpathy/micrograd | ★ 17.5k | Read `engine.py` (~100 lines); run `demo.ipynb` in Colab; then close it and rewrite from memory on Saturday |
| karpathy/nn-zero-to-hero | ★ 24.4k | Lecture notebooks under `lectures/micrograd`; follow along with the video |
| lightgbm-org/LightGBM | ★ 18.8k | Python quickstart regression; train on your unfolded features with batch-wise CV |
| rlabbe/Kalman-and-Bayesian-Filters-in-Python | ★ 19.3k | Colab: chapter 1 "The g-h filter". The intuition behind estimating a hidden variable from noisy sensors, which is what a soft sensor does |
| JustusvLiebig/Soft_Sensor_Experiments | ★ 103 (niche) | Browse the model zoo for industrial soft sensors; note which baselines they compare, to justify your LightGBM vs PLS choice |
| quarto-dev/quarto-cli | ★ 6.0k | `quarto create project manuscript`; build a PDF with one figure regenerated from code |

**NotebookLM kit · `LABLOOP-W08 · Technical report gate`** (gate week: add Video Overview + Slide Deck; then create `LABLOOP-ACT-II`)
- **Load (12):** (1) this chapter; (2) Luttmann 2012; (3) Kadlec 2009; (4) Karpathy micrograd video; (5) 3Blue1Brown backprop ch. 3; (6) Peyton Jones talk; (7) your technical report draft PDF; (8) F1 table; (9) soft-sensor error curve description; (10) `RELATED.md`; (11) week 4 notebook's saved Briefing doc; (12) P10 reviewer output.
- **Focus line:** `Focus: soft sensors as virtual instruments, how to write a technical report with an honest limitations section, and backpropagation as a chain of small local derivatives.`
- **Slide Deck prompt:** `Seven slides for a founder-scientist: the tank problem, IndPenSim, golden envelope and alarms (number), soft sensor (error at 50% of batch), FermLit F1 table (what LLMs can read), where it fails, the question I want to ask them.`
- **Custom question:** `Act as reviewer 2 on my technical report. Which claim is least supported by the figures, and which limitation did I forget?`

**Mentor check.** *"I have a DOI on a document that explains a method, a number and its failure modes, and five founders have it in their inbox."*

---


# ACT III · From sample to census (weeks 9-13)

> **Act III story.** A statistics office runs a pilot survey on 50 households, checks the questionnaire, trains the interviewers, and only then runs the national census. After the census, a second team re-interviews a sample to measure how consistent the counting was, and the office publishes the dataset with a methods note so anyone can use it. That is Act III. Your pilot was the gold set. The census is 2,000 papers. The second team is your second annotator. The publication is FermLit v0.5 on Hugging Face and Zenodo. And the first real finding, what fermentation papers forget to report, falls out of the census the way unemployment rates fall out of a labour survey.

---

### Week 9 · FermLit at scale

**Act III · Threads: B, M, FL · Next gate: 13**

**Welcome.** Scaling is where program managers shine and scientists often stumble. Going from 50 to 2,000 papers is not "run the loop 40 times"; it is a production run with a budget, a quality floor and a cost table. You set the floor in week 7 (the minimum F1 you accept); now you choose the cheapest backend that clears it, exactly like choosing a supplier who meets the spec at the lowest price. When the 2,000 JSONs land, you build the first view that matters: for each field, what percentage of papers even report it. This is the seed of Paper 1's finding. On Tuesday, you open the engineering side of biology: NPTEL's bioreactor lectures, where your week 1 equations reappear as formal mass balances (money in, money out, money accumulated, like a cash-flow statement for sugar). On Thursday, Karpathy's makemore: a tiny language model on organism names, so the machines reading your papers stop being magic.

**Where this week sits.**
- ← **From:** weeks 5-7 (corpus, schema, measured agent), week 1 (your mass balances).
- → **Feeds:** week 10 (descriptive analysis on 2,000 records), week 11 (reporting-completeness as a formal result), week 13 (dataset release), weeks 14-16 (Paper 1 Results).
- ⟳ **Returns in:** Paper 1 figure 1 (completeness by field), and in every LinkedIn post that says "X% of papers do not report pH".

**Goals.**
- 🎯 **This week:** 2,000 papers extracted within the F1 floor, a cost table, and a completeness table per field.
- 🏁 **Next milestone:** Gate 13 (FermLit v0.5 public, Paper 1 skeleton).
- 🌟 **Final:** the reporting-completeness finding that makes Paper 1 more than a dataset.

**The plan (v5, unchanged).**
- **Goal:** 2,000 papers extracted with the measured pipeline.
- **Tue (B9):** NPTEL bioreactors 1-3; balances in a notebook.
- **Thu (M9):** Karpathy makemore part 1; bigram on organism names for fun.
- **Build:** [ ] Mon: corpus to 2,000 papers (same query, more years, plus bioRxiv API for preprints). [ ] Wed: batch extraction with the cheapest backend that meets the F1 floor you set in week 7 (cost table with the v3 telemetry schema). Done when 2,000 JSONs exist. [ ] Fri: completeness dashboard: per field, % of papers that report it. This is the seed of the reporting finding.
- **Sun:** L-Sun: "2,000 fermentation papers in one table. X% do not report pH. Y% do not report volume." L-Wed.
- **Only one thing:** the completeness table.

**Learn it.**
- ▶ **Core**
  - Tue · NPTEL *Bioreactors* (Suraishkumar, IIT Madras), lectures 1-3 — nptel.ac.in/courses/102106053 (also on YouTube; ~75 min at 1.25x). Notebook: batch, fed-batch and continuous balances.
  - Thu · Karpathy, "The spelled-out intro to language modeling: building makemore" (YouTube, ~2 h; first 75 min Thursday).
- ＋ **Deeper**
  - MIT OCW 10.37 *Chemical and Biological Reaction Engineering* (Spring 2007), lecture notes on bioreactors — ocw.mit.edu/courses/10-37-chemical-and-biological-reaction-engineering-spring-2007.
  - Coursera *Industrial Biotechnology* (University of Manchester), module on fermentation processes.
  - LearnChemE "mass balance" screencasts (YouTube).
  - Karpathy "Intro to Large Language Models" (YouTube, 1 h) for a bird's-eye view.

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| karpathy/makemore | ★ 4.3k | Run `makemore.py` on a text file of 2,000 organism names extracted by FermLit; sample new "organisms" for fun |
| karpathy/nn-zero-to-hero | ★ 24.4k | `lectures/makemore` part 1 notebook in Colab |
| UKGovernmentBEIS/inspect_ai | ★ 2.8k | Read the "Tutorial" page; model your batch extraction as an eval task with logged cost and tokens (the Vara telemetry idea, done by a government AI institute) |
| ollama/ollama | ★ 180.8k | Benchmark tokens/second of your local Qwen 7B on 20 papers; add the row to the cost table |

**NotebookLM kit · `LABLOOP-W09 · FermLit scale-up`**
- **Load (9):** (1) this chapter; (2) NPTEL Bioreactors lectures 1-3 (YouTube links); (3) MIT 10.37 bioreactor notes PDF; (4) Karpathy makemore video; (5) your cost table; (6) completeness table; (7) week 7 F1 table; (8) Dagdelen 2024; (9) `notes/bio/B9.md`.
- **Focus line:** `Focus: scaling LLM extraction with a quality floor and a cost table, what the completeness of reporting per field reveals, and mass balances for batch, fed-batch and continuous bioreactors.`
- **Studio extra (Data Table):** `Per field: % of papers reporting it, likely reason it is missing, whether it matters for reproducing the experiment.`
- **Custom question:** `Which of my missing-field percentages could be an artifact of the extraction agent rather than of the papers? How would I test that?`

**Mentor check.** *"I can say what fraction of 2,000 fermentation papers report each key parameter, and why that number is trustworthy."*

---

### Week 10 · Inter-annotator agreement and the first real analysis

**Act III · Threads: B, M, FL · Next gate: 13**

**Welcome.** Two referees watching the same match will not call every foul the same way. If they disagree constantly, the rules are unclear, not the players. Cohen's kappa measures how much two annotators agree **beyond what chance alone would give**, and it is the number that tells reviewers whether your gold set is a standard or an opinion. This week you compute it per field, sit down with the disagreements, adjudicate them, and freeze gold v1. Then you do your first descriptive science: titer by organism and product, productivity versus scale, completeness by year. Descriptive is not a lesser kind of science; the first maps of any continent were descriptive, and they changed everything. On Tuesday, yields and kinetics on IndPenSim (how many grams of product per gram of sugar). On Thursday, Gaussian processes: a model that says not only "the titer will be 12 g/L" but "and I am unsure here, because I have never seen batches like this".

**Where this week sits.**
- ← **From:** week 6 gold set + second annotator; week 9 2,000 records.
- → **Feeds:** week 11 (formal completeness result), week 13 (Paper 1 skeleton figures), week 14 (Methods), week 16 (conformal prediction extends the GP's uncertainty idea), week 34 (judge-human agreement in labdocs).
- ⟳ **Returns in:** Paper 1 "Annotation agreement" paragraph and figures 1-4.

**Goals.**
- 🎯 **This week:** kappa per field in `eval/README.md`; gold v1 frozen; descriptive notebook; `make figs` regenerates figures 1-4.
- 🏁 **Next milestone:** Gate 13.
- 🌟 **Final:** the trust number in Paper 1's abstract.

**The plan (v5, unchanged).**
- **Goal:** kappa reported; first descriptive science.
- **Tue (B10):** NPTEL 4-6; yields on IndPenSim.
- **Thu (M10):** GPML ch. 1-2; GP soft sensor with bands.
- **Build:** [ ] Mon: compute Cohen's kappa per field between you and the second annotator; adjudicate disagreements; freeze gold v1. Done when kappa is in `eval/README.md`. [ ] Wed: descriptive analysis notebook: titer by organism and product, productivity vs scale, reporting completeness by year and journal. [ ] Fri: figures 1-4 of Paper 1 drafted (Plotly to static). Done when `make figs` regenerates them.
- **Sun:** L-Sun: "Two humans, 30 papers, kappa = X: how much can you trust a dataset built by an LLM?" L-Wed.
- **Only one thing:** kappa.

**Learn it.**
- ▶ **Core**
  - Tue · NPTEL *Bioreactors* lectures 4-6 (Monod, yields, maintenance) (75 min).
  - Thu · Rasmussen & Williams, *Gaussian Processes for Machine Learning*, ch. 1-2 — gaussianprocess.org/gpml (free) (45 min) **after** Görtler, Kehlbeck & Deussen (2019), "A Visual Exploration of Gaussian Processes", *Distill* — distill.pub/2019/visual-exploration-gaussian-processes (30 min, interactive; start here).
  - Build-block reading · McHugh (2012) kappa paper (carry over from week 6), results interpretation table.
- ＋ **Deeper**
  - ritvikmath "Gaussian Processes" (YouTube, whiteboard intuition).
  - StatQuest "Confidence Intervals, Clearly Explained" and "The Standard Error".
  - Krippendorff's alpha (Wikipedia page is good) for when you have more than two annotators or missing ratings.
  - *Modern Statistics for Modern Biology* (Holmes & Huber) ch. 3 "High quality graphics in R" — read for the design principles, ignore the R.

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| scikit-learn/scikit-learn | ★ 67.2k | `sklearn.metrics.cohen_kappa_score` per field; gallery example "Gaussian process regression: basic introductory example" |
| cornellius-gp/gpytorch | ★ 3.9k | "Simple GP Regression" tutorial notebook in Colab; add predictive variance bands to one soft-sensor feature |
| pln-fing-udelar/fast-krippendorff | ★ 161 | Compute alpha on the same annotations; compare with kappa and write one sentence on why they differ |
| plotly/plotly.py / mwaskom/seaborn | ★ 18.8k / ★ 14.0k | Build figure 1 (completeness by field) as a static, publication-ready image |

**NotebookLM kit · `LABLOOP-W10 · Agreement and analysis`**
- **Load (10):** (1) this chapter; (2) McHugh 2012; (3) Artstein & Poesio 2008; (4) Distill GP article (web); (5) GPML ch. 2 PDF; (6) NPTEL lectures 4-6; (7) your kappa table; (8) adjudication notes; (9) figures 1-4 captions; (10) week 9 completeness table.
- **Focus line:** `Focus: Cohen's kappa as referee agreement beyond chance, how to adjudicate annotation disagreements, descriptive analysis as map-making, and Gaussian processes as models that know when they do not know.`
- **Custom question:** `Interpret my per-field kappa values using the McHugh bands. Which fields should Paper 1 treat as reliable, cautious, or excluded?`

**Mentor check.** *"I can explain kappa to a founder with a referee analogy and tell them which FermLit fields are trustworthy."*

---

### Week 11 · Dataset engineering and the reporting finding

**Act III · Threads: B, M, FL · Next gate: 13**

**Welcome.** Anyone who has merged two companies' ERP systems knows the pain: one reports revenue in thousands, the other in units; one calls a customer "Bioceres S.A.", the other "BIOCERES". Data engineering is 80% of a merger and 80% of a dataset. This week FermLit gets its merger treatment: units normalised (g/L vs mg/L), organism names mapped to NCBI taxonomy IDs (so "*Y. lipolytica*" and "*Yarrowia lipolytica* Po1f" become the same species), duplicates removed. Then the finding becomes formal: which fields are missing, how that changed over twenty years, whether journals differ. And you draft the dataset card, the "nutrition label" that tells users what is inside and what is not. On Tuesday, oxygen transfer and scale-up: why a process that works in a 1-litre flask fails in a 10,000-litre tank (hint: it is harder to breathe in a crowded stadium than in a living room). On Thursday, Bayesian optimization: the art of choosing the next experiment when each one costs a week.

**Where this week sits.**
- ← **From:** weeks 9-10 (2,000 records, kappa, descriptive figures).
- → **Feeds:** week 13 (HF + Zenodo release), week 14 (Methods: normalisation), week 16 (bioRxiv data availability), week 23 (UniProt/taxonomy skills reused for enzymes), Half 2 Loop Audits ("what does your data look like after normalisation?").
- ⟳ **Returns in:** the dataset card that makes FermLit citable and reusable (L6).

**Goals.**
- 🎯 **This week:** normalised, deduplicated export; completeness-over-time result with a planned multiple-testing correction; draft HF dataset card with license.
- 🏁 **Next milestone:** Gate 13.
- 🌟 **Final:** a dataset others download and cite.

**The plan (v5, unchanged).**
- **Goal:** FermLit v0.5 as a proper dataset.
- **Tue (B11):** NPTEL kLa and scale-up; kLa estimate.
- **Thu (M11):** Frazier + BoTorch; BO on the simulator feed profile.
- **Build:** [ ] Mon: units normalization (g/L, mg/L, %), organism name normalization (NCBI taxonomy IDs via API), dedup. [ ] Wed: the reporting-completeness analysis as a formal result: which fields are missing, how that changed over 20 years, whether journals differ (with multiple-testing correction from B14 later). [ ] Fri: dataset card draft (HF), license (CC BY 4.0 for the extracted table; papers not redistributed), `dataset/` export.
- **Sun:** L-Sun: "The fermentation literature has a reporting problem, in numbers". L-Wed.
- **Only one thing:** normalized dataset export.

**Learn it.**
- ▶ **Core**
  - Tue · NPTEL *Bioreactors* lectures on oxygen transfer (kLa), mixing and scale-up (75 min).
  - Thu · Agnihotri & Batra (2020), "Exploring Bayesian Optimization", *Distill* — distill.pub/2020/bayesian-optimization (30 min, interactive; start here); then Frazier (2018), "A Tutorial on Bayesian Optimization", arXiv:1807.02811, sections 1-4 (45 min).
  - Build-block reading · Hugging Face "Dataset Cards" documentation (huggingface.co/docs/hub/datasets-cards); Wilkinson et al. (2016), "The FAIR Guiding Principles for scientific data management and stewardship", *Scientific Data* 3:160018 (4 pages).
- ＋ **Deeper**
  - Shields et al. (2021), "Bayesian reaction optimization as a tool for chemical synthesis", *Nature* 590:89-96 — BO beating expert chemists, the self-driving-lab argument.
  - Coursera *Research Data Management and Sharing* (UNC/Edinburgh), module on metadata.
  - NCBI Taxonomy and E-utilities documentation (ncbi.nlm.nih.gov/books/NBK25501) for the name-to-ID mapping.
  - Brunton "Bayesian optimization" lecture (search in his channel; verify).

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| meta-pytorch/botorch | ★ 3.6k | Tutorial `closed_loop_botorch_only` in Colab; replace the test function with your simulator's final titer as a function of 2 feed parameters |
| facebook/Ax | ★ 2.8k | "Getting started" BO loop in 10 lines; the higher-level alternative if BoTorch feels heavy |
| biopython/biopython | ★ 5.2k | `Bio.Entrez` to map 20 organism names to taxonomy IDs; cache results |
| huggingface/datasets (docs) | — | Create a private dataset repo, push a 100-row sample with a card; check the viewer renders |

**NotebookLM kit · `LABLOOP-W11 · Dataset engineering`**
- **Load (10):** (1) this chapter; (2) Wilkinson FAIR 2016; (3) HF dataset cards docs; (4) Gebru Datasheets (carry over); (5) Distill BO article; (6) Frazier 2018 PDF; (7) NPTEL kLa/scale-up lectures; (8) your normalisation rules as text; (9) completeness-over-time table; (10) dataset card draft.
- **Focus line:** `Focus: normalisation and deduplication as a data merger, FAIR principles and dataset cards, oxygen transfer as why scale-up fails, and Bayesian optimization as choosing the next expensive experiment.`
- **Custom question:** `Check my dataset card against FAIR and Datasheets for Datasets. List every question a reuser would ask that my card does not answer.`

**Mentor check.** *"I can explain why kLa limits scale-up, and I can show a dataset card that answers a stranger's questions before they ask."*

---

### Week 12 · Endorsement route opens; hybrid model v0

**Act III · Threads: B, M, BL, R · Next gate: 13**

**Welcome.** Two stories this week, one social and one technical, and they are the same story. The social one: arXiv, the main preprint server for ML, asks newcomers to be vouched for by someone already publishing there, like a private club where a member must sign your application. You open three routes at once (a founder, a cited author, a local academic), because one route is a hope and three are a plan. The technical one: your hybrid model. Imagine an experienced plant engineer (the Monod equations) working next to a sharp data analyst (a gradient-boosted model). The engineer gets the shape of the curve right from physics; the analyst learns only what the engineer gets wrong, the residual. With few batches, the engineer carries the team; with many, the analyst catches up. Measuring *when* that crossover happens is Paper 2. On Tuesday, the regulatory vocabulary of pharma (PAT, CPP, CQA); on Thursday, SINDy, an algorithm that rediscovers equations from data, so you can watch it find Monod in your simulator.

**Where this week sits.**
- ← **From:** week 1 Monod simulator (the engineer), week 8 LightGBM soft sensor (the analyst), week 4 batch-wise CV, week 8 technical report (proof of seriousness for endorsers).
- → **Feeds:** week 17 (full learning-curve experiment), week 18 (Paper 2 draft), weeks 19-20 (endorsement closes), week 27 (pharma vocabulary in the Loop Audit offer).
- ⟳ **Returns in:** Paper 2's central figure; the Part 9.5 routes.

**Goals.**
- 🎯 **This week:** three endorsement messages sent; hybrid model v0 compared to pure GBM on 10 batches; learning-curve harness plots for 5/10/20/50 batches × 5 seeds.
- 🏁 **Next milestone:** Gate 13 (one endorsement conversation alive).
- 🌟 **Final:** endorsed arXiv preprint (wk 20) and the question "when does biology help the model?" answered with numbers.

**The plan (v5, unchanged).**
- **Goal:** three routes to an arXiv endorser started; first hybrid model.
- **Tue (B12):** FDA PAT + ICH Q8; glossary.
- **Thu (M12):** Brunton ch. 7; SINDy recovers Monod from simulator data.
- **Build:** [ ] Mon: Part 9.5 routes: (1) message the two most-engaged founders from week 8 with the FermLit completeness figure and ask whether they would co-author or endorse; (2) email two authors you cite heavily (Prompt P13); (3) list three local academics (UBA, UNSAM, Leloir) who publish on arXiv q-bio. Done when three messages are sent. [ ] Wed: `batchlens/hybrid/`: Monod ODE with parameters fitted per batch + GBM on the residual; compare to pure GBM on 10 batches. [ ] Fri: learning-curve harness: train on 5, 10, 20, 50 batches, test on held-out normal and faulty batches, 5 seeds. Done when the curve plots.
- **Sun:** L-Sun: "Does knowing the biology help the model? Hybrid vs pure ML at 5, 10, 20, 50 batches (first look)". L-Wed.
- **Only one thing:** the three endorsement messages.

**Learn it.**
- ▶ **Core**
  - Tue · FDA (2004) *PAT — A Framework for Innovative Pharmaceutical Development, Manufacturing, and Quality Assurance* — fda.gov/media/71012/download (read sections I-IV, 45 min); ICH Q8(R2) sections 1-2 (ich.org) (15 min). Write the glossary.
  - Thu · Brunton & Kutz, *Data-Driven Science and Engineering* (2nd ed.), ch. 7 on SINDy — databookuw.com (free PDF) (40 min) + Steve Brunton "Sparse Identification of Nonlinear Dynamics (SINDy)" (YouTube, 30 min).
  - Build-block reading · Brunton, Proctor & Kutz (2016), "Discovering governing equations from data by sparse identification of nonlinear dynamical systems", *PNAS* 113(15):3932-3937 (abstract + figure 1).
- ＋ **Deeper**
  - Narayanan et al. (2020), "Bioprocessing in the digital age: the role of process models", *Biotechnology Journal* 15(1):e1900172.
  - Sansana et al. (2021), "Recent trends on hybrid modeling for Industry 4.0", *Computers & Chemical Engineering* 151:107365.
  - arXiv endorsement page — info.arxiv.org/help/endorsement.html (read once, carefully).
  - Steve Brunton "Physics Informed Machine Learning" series (YouTube), first 2 videos.

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| dynamicslab/pysindy | ★ 1.9k | Colab: `examples/tutorial_1`; then feed it X and S from your simulator and see whether it recovers a Monod-like term |
| dynamicslab/databook_python | ★ 1.0k | The book's Python notebooks for chapter 7 (SINDy) in Colab |
| jrcramos/Hybrid-modeling-of-bioreactor-with-LSTM | ★ 24 (niche, academic) | Read the README and model structure; a published example of "mechanistic + neural" for cell culture to compare against your "Monod + GBM" |
| casadi/casadi | ★ 2.3k | Optional: fit Monod parameters per batch with an optimizer; only if `scipy.optimize` struggles |

**NotebookLM kit · `LABLOOP-W12 · Endorsement and hybrid model`**
- **Load (11):** (1) this chapter; (2) FDA PAT 2004 PDF; (3) ICH Q8(R2) PDF; (4) Brunton & Kutz ch. 7 PDF; (5) SINDy YouTube; (6) Brunton et al. 2016 PNAS; (7) Narayanan 2020; (8) von Stosch et al. 2014 (if accessible; else abstract); (9) arXiv endorsement page; (10) your hybrid v0 vs GBM table; (11) your three endorsement messages.
- **Focus line:** `Focus: hybrid modeling as an engineer and an analyst working together, SINDy as equation discovery, the pharma vocabulary of PAT, CPP and CQA, and how arXiv endorsement works.`
- **Custom question:** `Given my hybrid v0 results on 10 batches, predict where the crossover with pure ML should appear on the learning curve, and what result would falsify my hypothesis.`

**Mentor check.** *"I can draw the hybrid architecture on a napkin, say what PAT means to a QA manager, and name my three endorsement routes."*

---

### Week 13 · GATE 13: FermLit v0.5 public, Paper 1 outlined

**Act III (finale) · Threads: B, M, FL, R · GATE 13**

**Welcome.** A census office's proudest day is not when the counting ends; it is when the tables go public and a journalist, a mayor and a researcher all use them for different purposes. Today FermLit goes public: dataset on Hugging Face with its card, Zenodo DOI, `v0.5` release. Then you do something that feels premature and is actually the professional move: you build the skeleton of Paper 1 before writing a single paragraph of prose. Title with a number. Abstract with numbers. Six figures in their slots. Methods as bullets. It is the architecture drawing before construction, the part you have done a thousand times as a program manager: scope, structure, dependencies. Then Reviewer 2 (Prompt P10) attacks it, so the holes appear while they are cheap. On Tuesday, statistics for biologists, and why they distrust p-values from n = 3. On Thursday, the hybrid modeling review and Neural ODEs, the modern cousin of what you built last week.

**Where this week sits.**
- ← **From:** Act III weeks 9-12.
- → **Feeds:** weeks 14-16 (write, review, submit Paper 1), week 16 (founders and funds receive a real link), week 19 (personal site), week 36 (journal version).
- ⟳ **Returns in:** L6 "used" when someone cites FermLit.

**Goals.**
- 🎯 **This week:** HF dataset + card + Zenodo DOI; Paper 1 skeleton with every figure; P10 list of unsupported claims.
- 🏁 **This gate (13):** dataset public, skeleton complete, one endorsement conversation alive.
- 🌟 **Final:** Paper 1 and a dataset with a life beyond you.

**The plan (v5, unchanged).**
- **Goal:** dataset on Hugging Face and Zenodo; Paper 1 skeleton with all figures.
- **Tue (B13):** MSMB ch. 1-2.
- **Thu (M13):** von Stosch review + Neural ODE abstract.
- **Build:** [ ] Mon: HF dataset upload with card, Zenodo DOI, `fermlit v0.5` release. [ ] Wed: Paper 1 skeleton with Prompt P9: title, abstract with numbers, 6 figures placed, methods bullet-pointed. [ ] Fri: Prompt P10 (reviewer 2) on the skeleton; list of every claim without evidence; plan the fixes.
- **Sun:** **M3** deep dive: "FermLit: 2,000 fermentation papers, one table, and what is missing from them". **Z3**, **HF**. L-Sun. Gauge: L2 solid, L3 in sight.
- **Gate 13 passes when:** dataset public with DOI and card; Paper 1 skeleton has every figure; at least one endorsement conversation alive.
- **Only one thing:** the HF dataset.

**Learn it.**
- ▶ **Core**
  - Tue · Holmes & Huber, *Modern Statistics for Modern Biology*, ch. 1 (generative models for discrete data) and ch. 2 (statistical modeling) — search "Modern Statistics for Modern Biology Holmes Huber" (free online book; EMBL link may move) (75 min).
  - Thu · von Stosch et al. (2014), "Hybrid semi-parametric modeling in process systems engineering: past, present and future", *Computers & Chemical Engineering* 60:86-101 (sections 1-3); Chen et al. (2018), "Neural Ordinary Differential Equations", arXiv:1806.07366 (abstract + figure 1).
  - Build-block reading · Mensh & Kording (2017), "Ten simple rules for structuring papers", *PLoS Computational Biology* 13(9):e1005619.
- ＋ **Deeper**
  - Steve Brunton "Neural ODEs" (YouTube).
  - Coursera *Writing in the Sciences* (Stanford, Kristin Sainani), module 1 "Principles of effective writing" — preview for week 14.
  - Karniadakis et al. (2021), "Physics-informed machine learning", *Nature Reviews Physics* 3:422-440 (figures 1-2).
  - Stanford CS197 (cs197.stanford.edu), lecture on "vertical paper structure" (verify lecture title).

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| rtqichen/torchdiffeq | ★ 6.5k | `examples/ode_demo.py`: learn a spiral's dynamics; read how a neural network replaces the right-hand side of an ODE |
| patrick-kidger/diffrax | ★ 2.1k | Optional JAX alternative with excellent docs on neural ODEs |
| quarto-dev/quarto-cli | ★ 6.0k | Put the Paper 1 skeleton into a Quarto manuscript; `make paper` builds a PDF with placeholder figures |
| huggingface/notebooks | ★ 4.6k | Browse the datasets examples for pushing a dataset with a card from code |

**NotebookLM kit · `LABLOOP-W13 · FermLit release and Paper 1 outline`** (gate week: add Video Overview + Slide Deck; then create `LABLOOP-ACT-III`)
- **Load (12):** (1) this chapter; (2) Mensh & Kording 2017; (3) MSMB ch. 1-2; (4) von Stosch 2014; (5) Chen 2018 PDF; (6) your HF dataset page; (7) dataset card; (8) Paper 1 skeleton; (9) P10 output; (10) figures 1-6 captions; (11) `RELATED.md`; (12) `LABLOOP-ACT-II` Briefing doc.
- **Focus line:** `Focus: how a dataset paper is structured, what makes FermLit's contribution novel, and the gap between descriptive findings and claims.`
- **Slide Deck prompt:** `Eight slides presenting FermLit to a fermentation scientist: why literature is data, corpus, schema, gold set and kappa, extraction F1, completeness finding, how to use the dataset, limitations.`
- **Custom question:** `Using Mensh & Kording's rules, rewrite my abstract skeleton so that each sentence carries one number and the gap is stated in the second sentence.`

**Mentor check.** *"FermLit has a DOI and a card, and I can pitch Paper 1 in three sentences with three numbers."*

---

# ACT IV · Becoming an author (weeks 14-20)

> **Act IV story.** For 25 years you wrote status reports read by people who already trusted you. Now you write for strangers who are paid to doubt you. It is the difference between an internal memo and a listed company's audited annual report: every number footnoted, every method disclosed, every risk stated, signed by an auditor. Paper 1 is your first annual report (bioRxiv, week 16). Paper 2 is the second, with an outside signature (the arXiv endorser, week 20). By the end of the act you are no longer "an engineer learning biology"; you are an author.

---

### Week 14 · Write Paper 1

**Act IV · Threads: B, M, FL · Next gate: 16**

**Welcome.** Writers have a saying: you do not write a book, you write a page a day and one day it is a book. This week you write three sections in three build blocks, and the order matters. Monday: Introduction and Related work, the "why should anyone care, and what exists already". Wednesday: Methods, the recipe precise enough that another cook produces the same dish. Friday: Results and Limitations, the plates on the table and an honest note about what burned. Methods is the section to love. For a reviewer, Methods is where trust is won or lost, and it is the section where your program-manager brain has an unfair advantage: it is a process description with acceptance criteria. On Tuesday, multiple testing: if you compare 20 journals, one will look "significantly different" by pure chance, like finding a "lucky" roulette table after watching twenty. On Thursday, calibrated judges for fuzzy fields ("*Y. lipolytica* Po1f" vs "*Yarrowia lipolytica*").

**Where this week sits.**
- ← **From:** week 13 skeleton and reviewer list; weeks 5-11 artifacts.
- → **Feeds:** week 15 (review and human read), week 16 (submission), week 31 (Paper 3 uses the same writing process), week 36 (journal version).
- ⟳ **Returns in:** every future paper: your Methods style becomes your signature.

**Goals.**
- 🎯 **This week:** full draft of Paper 1 (Intro, Related work, Methods, Data availability, Results, Limitations); multiple-testing correction applied; calibrated judge for fuzzy fields.
- 🏁 **Next milestone:** Gate 16 (bioRxiv DOI).
- 🌟 **Final:** L3 preprinted, the rung that makes founders answer.

**The plan (v5, unchanged).**
- **Goal:** full draft.
- **Tue (B14):** MSMB ch. 6; correction applied to the journal comparison.
- **Thu (M14):** CS336 L12 + CS224N judge lecture; calibrated judge for fuzzy fields (does "Yarrowia lipolytica Po1f" match "Y. lipolytica"?).
- **Build:** [ ] Mon: Introduction and Related work (from `RELATED.md`, every citation verified via Semantic Scholar). [ ] Wed: Methods (corpus, schema, guidelines, agents, evaluation, kappa) and Data availability. [ ] Fri: Results (completeness finding, per-field F1, descriptive tables) and Limitations (honest: extraction errors, OA bias, English-only).
- **Sun:** L-Sun: "Writing my first paper at 25 years of career: what a methods section really is". L-Wed.
- **Only one thing:** Methods section.

**Learn it.**
- ▶ **Core**
  - Tue · *Modern Statistics for Modern Biology* ch. 6 "Testing" (multiple testing, FDR) (60 min); StatQuest "False Discovery Rates, FDR, clearly explained" and "p-hacking: What it is and how to avoid it!" (YouTube, 20 min).
  - Thu · Stanford CS336 lecture on evaluation (cs336.stanford.edu, videos on Stanford Online YouTube) (60 min); your week 7 notes on Zheng et al. 2023 judge biases (15 min).
  - Writing · Coursera *Writing in the Sciences* (Stanford), modules 1-2 (cutting clutter, active voice) — watch at 1.5x during Saturday buffer if needed.
- ＋ **Deeper**
  - Gopen & Swan (1990), "The Science of Scientific Writing", *American Scientist* 78:550-558. The best 9 pages ever written on why readers get lost.
  - Whitesides (2004), "Whitesides' Group: Writing a Paper", *Advanced Materials* 16(15):1375-1377. Outline-first writing from a legendary chemist.
  - Benjamini & Hochberg (1995), "Controlling the false discovery rate", *JRSS B* 57(1):289-300 (abstract + procedure).
  - *Scientific Data* "Data Descriptor" article guidelines (nature.com/sdata) — preview of the week 36 venue.

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| statsmodels/statsmodels | ★ 11.6k | `statsmodels.stats.multitest.multipletests` with `fdr_bh` on the journal comparison; report adjusted p-values |
| UKGovernmentBEIS/inspect_ai | ★ 2.8k | Implement the fuzzy-field judge as a scorer; log judge vs human agreement on 50 pairs |
| EleutherAI/lm-evaluation-harness | ★ 14.0k | Read how tasks define metrics and few-shot prompts; borrow the discipline, not the tool |
| Future-House/paper-qa | ★ 9.2k | Point it at your `RELATED.md` PDFs and ask "which of these papers report inter-annotator agreement?" to double-check your related-work claims (verify every answer by opening the paper) |

**NotebookLM kit · `LABLOOP-W14 · Paper 1 full draft`**
- **Load (11):** (1) this chapter; (2) Gopen & Swan 1990; (3) Whitesides 2004; (4) Mensh & Kording 2017; (5) MSMB ch. 6; (6) StatQuest FDR; (7) CS336 evaluation lecture; (8) Paper 1 draft (PDF or Doc); (9) `RELATED.md`; (10) kappa + F1 tables; (11) `LABLOOP-THESIS` papers you cite (add to the thesis notebook, not here, if over 15).
- **Focus line:** `Focus: how to write the Methods of a dataset-and-method paper so it is reproducible, multiple-testing correction for many comparisons, and calibrated LLM judges.`
- **Custom question:** `Using Gopen & Swan, find the five sentences in my Methods where the reader's expectation is violated (topic and stress positions), and rewrite them.`

**Mentor check.** *"My Methods section lets a stranger rebuild FermLit, and I can explain why I corrected for multiple testing."*

---

### Week 15 · Review, revise, human read

**Act IV · Threads: B, M, FL, R · Next gate: 16**

**Welcome.** Before a bridge opens, two different inspectors walk it: one checks the calculations, the other checks the welds. This week your paper gets two inspectors. The first is Reviewer 2, the famously harsh anonymous reviewer, played by an LLM twice. The second is a human scientist, the one inspection that cannot be automated. You will ask for it, and you may pay a postdoc for two hours. That is not a weakness: professionals pay for audits. On Tuesday, you practise the skill of reading scientific papers fast, with a three-pass method: first the "bird's-eye" pass (5 minutes), then the "grasp the content" pass (1 hour), then the "rebuild it in your head" pass (only for the few papers that matter). On Thursday, information extraction as a field and inter-annotator agreement written properly into Methods.

**Where this week sits.**
- ← **From:** week 14 draft.
- → **Feeds:** week 16 submission; week 19 (same human reviewer for Paper 2); weeks 31-32 (co-author review loop for Paper 3).
- ⟳ **Returns in:** your reputation among the people who read drafts; first reviewers often become endorsers or co-authors.

**Goals.**
- 🎯 **This week:** every [VERIFY] fixed; human review requested (and ideally received); `make paper` builds the final PDF.
- 🏁 **Next milestone:** Gate 16.
- 🌟 **Final:** a network of scientists who have read your work before it was public.

**The plan (v5, unchanged).**
- **Goal:** a draft a scientist has read.
- **Tue (B15):** three papers with P11, IMRaD noted.
- **Thu (M15):** Jurafsky & Martin IE chapter; kappa written into Methods.
- **Build:** [ ] Mon: Prompt P10 twice; fix every [VERIFY]. [ ] Wed: send to one human scientist (co-author candidate, endorser, or paid postdoc, USD ~100 for 2 hours). [ ] Fri: figures final; `make paper` builds the PDF from the repo.
- **Sun:** L-Sun: "Reviewer 2 is an LLM and a postdoc; both found things". L-Wed.
- **Only one thing:** human review requested.

**Learn it.**
- ▶ **Core**
  - Tue · S. Keshav (2007), "How to Read a Paper", *ACM SIGCOMM Computer Communication Review* 37(3):83-84 (2 pages, 15 min); then three FermLit papers with P11 and IMRaD noted (60 min).
  - Thu · Jurafsky & Martin, *Speech and Language Processing* (3rd ed. draft, web.stanford.edu/~jurafsky/slp3), chapter on information extraction (relation extraction and evaluation sections) (60 min); write the kappa paragraph into Methods.
- ＋ **Deeper**
  - Coursera *Writing in the Sciences*, module on "The format of a scientific manuscript" and "Peer review".
  - Bourne (2005), "Ten simple rules for getting published", *PLoS Computational Biology* 1(5):e57.
  - ASAPbio preprint FAQ (asapbio.org) — what preprints mean for credit and journals.
  - StatQuest "Statistical Power, Clearly Explained" — reviewers may ask whether 50 gold papers is enough.

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| allenai/scispacy | ★ 2.0k | Run a biomedical NER model on 3 FermLit papers; compare what it finds (organisms, chemicals) with your agent's extraction. A classic baseline reviewers may ask about |
| grobidOrg/grobid | ★ 5.1k | Parse one PDF-only paper into TEI XML; note how references and sections are recovered. Useful for the non-XML tail of the corpus |
| docling-project/docling | ★ 66.3k | Convert the same PDF with Docling; compare tables extraction with GROBID |
| quarto-dev/quarto-cli | ★ 6.0k | Final `make paper`; check figures regenerate from a clean clone |

**NotebookLM kit · `LABLOOP-W15 · Paper 1 review and revision`**
- **Load (10):** (1) this chapter; (2) Keshav 2007; (3) Jurafsky & Martin IE chapter PDF; (4) Bourne 2005; (5) Paper 1 draft v2; (6) P10 outputs (both rounds); (7) human reviewer's comments; (8) three P11 digests; (9) kappa paragraph; (10) `QUESTIONS.md`.
- **Focus line:** `Focus: how peer review works, how to respond to reviewer criticism without defensiveness, the three-pass way to read papers, and how information extraction is evaluated.`
- **Custom question:** `Merge the LLM reviewer's and the human reviewer's comments into one response table: comment, severity, my change, where in the paper.`

**Mentor check.** *"A human scientist has read my paper, and I can read any paper's core in 10 minutes."*

---

### Week 16 · GATE 16: Paper 1 on bioRxiv (L3)

**Act IV · Threads: B, M, FL, EL (literacy), R · GATE 16**

**Welcome.** There is a moment in every explorer's story when the map leaves their hands. Before it, it is a draft; after it, it belongs to everyone who wants to check it. This week you press "submit" on bioRxiv. It will feel anticlimactic (a form, a PDF, a category) and then, a few days later, a DOI appears, and your name is attached to a piece of science forever. You also translate it for humans: a LinkedIn article for non-scientists, and a quiet email to the founders and companies you admire, with no ask. Meanwhile Act V is already knocking: on Tuesday you meet protein language models, the idea that the "language" of amino acids can be learned the way LLMs learned English, and you run ESM-2 on your own 3050. On Thursday, conformal prediction: honest intervals with a guarantee, like a weather forecast that promises "90% of my 90% intervals will contain tomorrow's temperature" and keeps the promise.

**Where this week sits.**
- ← **From:** weeks 5-15 (everything FermLit).
- → **Feeds:** week 17 (conformal intervals in Paper 2), weeks 17, 21-23 (ESM-2 is the engine of Paper 3), week 19 (applications and site), week 27 (offer built on preprints), week 36 (journal submission).
- ⟳ **Returns in:** every message you send from now on; the preprint is your business card.

**Goals.**
- 🎯 **This week:** bioRxiv submitted with code and data DOIs; LinkedIn article; preprint sent to 10 founders and 3 companies; ESM-2 embedding 100 sequences; conformal intervals on the soft sensor.
- 🏁 **This gate (16):** L3 preprinted.
- 🌟 **Final:** the credential that turns "who is this?" into "I read your preprint".

**The plan (v5, unchanged).**
- **Goal:** submitted.
- **Tue (B16):** ESM-2 paper + MIT 6.874 protein lecture; ESM-2 running.
- **Thu (M16):** conformal prediction sections 1-2; intervals on the soft sensor.
- **Build:** [ ] Mon: incorporate human review. [ ] Wed: bioRxiv submission (category: Bioengineering or Systems Biology), code and data DOIs in the paper. [ ] Fri: LinkedIn article version of the paper for non-scientists; send the preprint to the 10 founders you have talked to and the 3 scientific-data companies you admire (no ask, just the link).
- **Sun:** L-Sun: the preprint. **X1**. Gauge: **L3**.
- **Gate 16 passes when:** bioRxiv shows the DOI.
- **Only one thing:** the submission.

**Learn it.**
- ▶ **Core**
  - Tue · Lin et al. (2023), "Evolutionary-scale prediction of atomic-level protein structure with a language model", *Science* 379(6637):1123-1130 — abstract, introduction and figures (40 min); MIT 6.874 *Deep Learning in the Life Sciences* (Manolis Kellis, 2021, YouTube playlist), protein lecture (first 40 min).
  - Thu · Angelopoulos & Bates (2021), "A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification", arXiv:2107.07511, sections 1-2 (60 min).
  - Build-block reading · bioRxiv submission guide (biorxiv.org/submit-a-manuscript).
- ＋ **Deeper**
  - Rives et al. (2021), "Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences", *PNAS* 118(15) — the ESM-1b paper, for the "why it works" story.
  - DeepMind, "AlphaFold: The making of a scientific breakthrough" (YouTube, 8 min) — the emotional context of protein AI.
  - Nobel Prize in Chemistry 2024 lectures (Baker, Hassabis, Jumper) — nobelprize.org.
  - Anastasios Angelopoulos' conformal prediction tutorial videos (YouTube; verify).
  - Valence Labs (YouTube), search "ESM" or "protein language models" in the M2D2 reading group.

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| huggingface/notebooks | ★ 4.6k | Colab: `examples/protein_language_modeling.ipynb` — fine-tune/embed with ESM-2 through `transformers`; switch the model to `facebook/esm2_t33_650M_UR50D` in fp16 on your 3050 |
| facebookresearch/esm | ★ 4.2k (archived: read, do not depend) | Read the README's "ESM-2" and embedding extraction sections; the original reference |
| evolutionaryscale/esm (now redirects to Biohub/esm) | ★ 3.0k | Read only; the newer ESM C/ESM3 family. Parked for Half 2 |
| aangelopoulos/conformal-prediction | ★ 1.1k | Colab: `notebooks/meps-cqr.ipynb` (conformalized quantile regression on tabular data); map it onto your soft sensor |
| scikit-learn-contrib/MAPIE | ★ 1.6k | Regression example with `SplitConformalRegressor` (or the current API); get coverage on held-out batches in 20 lines |

**NotebookLM kit · `LABLOOP-W16 · Paper 1 preprint gate`** (gate week: add Video Overview + Slide Deck)
- **Load (11):** (1) this chapter; (2) your bioRxiv PDF; (3) LinkedIn article draft; (4) Lin et al. 2023 (abstract page + PDF if accessible); (5) MIT 6.874 protein lecture; (6) Angelopoulos & Bates PDF; (7) HF protein notebook GitHub page; (8) MAPIE docs page; (9) your ESM-2 embedding log; (10) conformal coverage result; (11) week 13 Slide Deck.
- **Focus line:** `Focus: what a bioRxiv preprint is and what it changes, protein language models as learning the grammar of proteins, and conformal prediction as a forecast that keeps its promise.`
- **Slide Deck prompt:** `Six slides for a non-scientist founder audience: the reporting problem in fermentation literature, what FermLit is, the one figure, what it means for their lab records, how to use the dataset, link.`
- **Custom question:** `Explain to me like a curious engineer why an embedding from ESM-2 might carry information about thermostability, and what could make that signal spurious.`

**Mentor check.** *"I am the author of a preprint with a DOI. I can explain conformal coverage and what a protein embedding is."*

---


### Week 17 · Paper 2 experiments, complete

**Act IV · Threads: B, M, BL, EL (literacy) · Next gate: 20**

**Welcome.** A clinical trial is not "we gave the drug to some people and they felt better". It is arms, doses, randomisation, repeats, and a pre-declared outcome. This week Paper 2 becomes a proper trial. Three arms (mechanistic, pure ML, hybrid), four "doses" of data (5, 10, 20, 50 batches), five seeds each so luck cannot fake a result, conformal intervals so every number carries its honest uncertainty, and a stress test on faulty batches to see which model notices trouble first. The result will be a learning curve, which is simply a chart of "how good is each model as a function of how much data it saw". If the hybrid wins with few batches and loses with many, that is exactly the story a data-poor biotech startup needs to hear. On Tuesday, the protein benchmarks FLIP and Meltome, the "exam papers" for protein models. On Thursday, you embed Meltome sequences with ESM-2 and fit a simple ridge head: your first protein baseline.

**Where this week sits.**
- ← **From:** week 12 harness and hybrid v0, week 16 conformal intervals, week 4 batch-wise CV, week 3 hours-to-detection.
- → **Feeds:** week 18 (Paper 2 draft), week 20 (arXiv), weeks 21-22 (the Meltome baseline you start today becomes Paper 3's baseline table).
- ⟳ **Returns in:** Paper 2 figures; JOSS (wk 36) as the showcase usage of batchlens.

**Goals.**
- 🎯 **This week:** full learning-curve experiment with seeds and coverage; transfer to faulty batches; `paper-2` repo with `make figs`; `batchlens v0.3` + Z4; ESM-2 + ridge baseline number on Meltome.
- 🏁 **Next milestone:** Gate 20 (Paper 2 on arXiv).
- 🌟 **Final:** the result that answers "when does biology help the model?".

**The plan (v5, unchanged).**
- **Goal:** every number in Paper 2 exists.
- **Tue (B17):** FLIP + Meltome papers; Meltome split loaded.
- **Thu (M17):** ESM-2 embeddings + ridge head on Meltome; baseline number.
- **Build:** [ ] Mon: full learning-curve experiment (mechanistic, pure ML, hybrid × 5/10/20/50 batches × 5 seeds), conformal intervals, coverage. [ ] Wed: transfer to faulty batches; hours-to-detection with each model. [ ] Fri: `paper-2` repo with `make figs`; `batchlens v0.3` + **Z4**.
- **Sun:** L-Sun: "Hybrid wins below 20 batches, loses above 50: the full curve". L-Wed.
- **Only one thing:** the learning curves with seeds.

**Learn it.**
- ▶ **Core**
  - Tue · Dallago et al. (2021), "FLIP: Benchmark tasks in fitness landscape inference for proteins", NeurIPS Datasets and Benchmarks (bioRxiv 2021.11.09.467890) — introduction, task table, thermostability split (40 min); Jarzab et al. (2020), "Meltome atlas — thermal proteome stability across the tree of life", *Nature Methods* 17:495-503 — abstract and figure 1 (20 min).
  - Thu · Hugging Face model card `facebook/esm2_t33_650M_UR50D` (10 min) + MIT 6.874 protein lecture second half (40 min); code block with ridge head.
  - StatQuest "Ridge vs Lasso Regression, Visualized!!!" (YouTube) for the head you fit.
- ＋ **Deeper**
  - Yang, Wu & Arnold (2019), "Machine-learning-guided directed evolution for protein engineering", *Nature Methods* 16:687-694. The classic review; read it before week 21.
  - Frances Arnold, Nobel Lecture 2018 "Innovation by Evolution: Bringing New Chemistry to Life" (nobelprize.org).
  - Notin et al. (2023), "ProteinGym: large-scale benchmarks for protein fitness prediction and design", NeurIPS Datasets and Benchmarks.
  - Andrew White, *Deep Learning for Molecules and Materials* — dmol.pub, chapter on "Deep learning on sequences" (with Colab buttons).

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| J-SNACKKB/FLIP | ★ 141 (niche; official benchmark by the paper authors) | Download the `meltome` splits; read the README on "mixed", "human" and "human_cell" splits; load them with pandas |
| OATML-Markslab/ProteinGym | ★ 470 (niche; the reference fitness benchmark) | Browse the leaderboard page and the baselines list; note where ESM-2 sits |
| huggingface/notebooks | ★ 4.6k | Reuse `protein_language_modeling.ipynb` to cache mean-pooled embeddings to disk |
| yangkky/Machine-learning-for-proteins | ★ 1.7k | The curated reading list by Kevin Yang; pick 2 papers for `PARKING.md`, not more |
| scikit-learn/scikit-learn | ★ 67.2k | `RidgeCV` on embeddings → melting temperature; report Spearman correlation with a bootstrap interval |

**NotebookLM kit · `LABLOOP-W17 · Paper 2 experiments`**
- **Load (11):** (1) this chapter; (2) FLIP preprint PDF; (3) Meltome atlas abstract page; (4) ESM-2 HF model card; (5) Yang, Wu & Arnold 2019; (6) Angelopoulos & Bates (carry over); (7) your learning-curve figure description; (8) coverage table; (9) hours-to-detection per model; (10) Meltome baseline result; (11) `LABLOOP-W12` Briefing doc.
- **Focus line:** `Focus: learning curves as a data-dose trial for three model types, why seeds and intervals matter, and protein thermostability benchmarks (FLIP, Meltome) as exam papers for protein models.`
- **Studio extra (Infographic):** `Learning curve explainer: x-axis number of batches, three lines (mechanistic, ML, hybrid), what the crossover means for a startup with 8 batches.`
- **Custom question:** `Where could leakage hide in my learning-curve experiment and in my Meltome split? List concrete checks.`

**Mentor check.** *"I can show the learning curve and say, with an interval, below how many batches the hybrid model is the better bet."*

---

### Week 18 · Write Paper 2

**Act IV · Threads: B, M, BL, R · Next gate: 20**

**Welcome.** The second time you do anything hard, it is a different activity. Paper 1 taught you the shape of a paper; Paper 2 is where you learn the voice. A modeling paper is read by sceptical engineers who have seen a hundred "our model beats the baseline" claims, so the voice must be calm, precise and generous with limitations: simulated data, one process, no real plant. That honesty is not a weakness; it is what makes the one strong claim believable. Think of a good investment memo: the best ones spend a page on the risks, and that is why the partner signs. This week you also do some quiet scouting for Act V: on Tuesday you read two papers from extremophile founders (organisms that live in hot springs, salt flats or Antarctic ice), because their enzymes are the ones Paper 3 will study. On Thursday, a legendary talk on how to write a great research paper and the NeurIPS checklist, which is basically a pre-flight checklist for claims.

**Where this week sits.**
- ← **From:** week 17 experiments and figures; week 14-15 writing process.
- → **Feeds:** week 19 (review), week 20 (submission), week 21 (the extremophile papers become partner-call preparation), week 36 (journal version).
- ⟳ **Returns in:** Medium deep dive M4 and the week 38 talk.

**Goals.**
- 🎯 **This week:** Paper 2 full draft (Intro, Related work, Methods, Results, Limitations, Reproducibility statement); M4 deep dive; two founder-paper digests with outreach lines.
- 🏁 **Next milestone:** Gate 20.
- 🌟 **Final:** your second preprint, and the start of the partner conversations that make Paper 3 possible.

**The plan (v5, unchanged).**
- **Goal:** full draft.
- **Tue (B18):** two extremophile founder papers; outreach lines.
- **Thu (M18):** Peyton Jones talk + NeurIPS checklist; rewrite.
- **Build:** [ ] Mon: Intro + Related work (hybrid modeling, soft sensors, IndPenSim benchmarks; all verified). [ ] Wed: Methods + Results. [ ] Fri: Limitations (simulated data, one process, no real plant), Reproducibility statement.
- **Sun:** **M4** deep dive: "When does biology help the model? A data-scarcity study on 100 open batches". L-Sun.
- **Only one thing:** Results section.

**Learn it.**
- ▶ **Core**
  - Tue · Two founder or scientific-lead papers from extremophile-focused GridX companies (Puna Bio, Antarka, APEXzymes authors), found via Europe PMC / Google Scholar, digested with P11 (75 min). Background first: Crash Course Biology "Old & Odd: Archaea, Bacteria & Protists" (#35, YouTube, 12 min).
  - Thu · Simon Peyton Jones, "How to write a great research paper" (YouTube, Microsoft Research, ~1 h) + NeurIPS Paper Checklist (neurips.cc, current year's author guidelines) (30 min).
- ＋ **Deeper**
  - Coursera *Writing in the Sciences* (Stanford), module "Writing the results and discussion".
  - Kapoor & Narayanan (2023) leakage paper (carry over) — use its checklist on Paper 2's Methods.
  - Rothschild & Mancinelli (2001), "Life in extreme environments", *Nature* 409:1092-1101 — the classic review of extremophiles.
  - Stanford CS197, lecture on "writing the paper" (verify title).

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| quarto-dev/quarto-cli | ★ 6.0k | Paper 2 manuscript project; cross-references for figures and equations |
| lukasmasuch/best-of-ml-python | ★ 23.8k | Check the "interpretability" and "time series" sections for any widely used soft-sensor baseline you forgot to cite |
| google-research/tuning_playbook | ★ 30.3k | Re-read "scientific vs nuisance hyperparameters" and write two sentences in Methods on how you tuned fairly across arms |
| paper-2-hybrid-softsensor (your repo) | — | Clean clone → `make figs` → compare hashes with your local figures |

**NotebookLM kit · `LABLOOP-W18 · Paper 2 full draft`**
- **Load (10):** (1) this chapter; (2) Paper 2 draft; (3) Peyton Jones talk; (4) NeurIPS checklist page; (5) von Stosch 2014 + Sansana 2021 (hybrid related work); (6) Luttmann 2012 + Kadlec 2009 (soft-sensor related work); (7) Goldrick 2019; (8) two extremophile founder papers; (9) P10 output; (10) M4 draft.
- **Focus line:** `Focus: writing a modeling paper with honest limitations, the NeurIPS checklist as a pre-flight checklist for claims, and what extremophiles are and why their enzymes matter to industry.`
- **Custom question:** `Run the NeurIPS checklist against my Paper 2 draft: for each item, answer yes/no/partial and quote the sentence that justifies it.`

**Mentor check.** *"My Paper 2 draft makes one strong, bounded claim and lists its limitations before a reviewer does."*

---

### Week 19 · Endorsement closed; review; site

**Act IV · Threads: B, M, BL, R · Next gate: 20**

**Welcome.** Around this point in an expedition, the team walks into a town for the first time in months. People want to know who they are. You need three things: a letter of introduction (the endorser's yes), a second pair of eyes on the report (review), and a sign on the door (a one-screen site). You also start knocking on doors that were closed at week 0: five applications to scientific-data and bioprocess-AI companies with your two preprints as the cover letter. This is where the employee track and the consulting track stop being theory. On Tuesday you get omics literacy through the most famous single-cell tutorial in the world (PBMC 3k with scanpy), so a single-cell talk no longer sounds like a foreign language. On Thursday you make every number in both papers carry an interval, because a number without an interval, to a scientist, is a rumour.

**Where this week sits.**
- ← **From:** week 12 endorsement routes, week 15 reviewer, weeks 16-18 preprints and drafts.
- → **Feeds:** week 20 (submission), week 27 (offer page links to the site), week 32 (second round of applications), week 37 (interviews).
- ⟳ **Returns in:** every gate, when the site and headline update.

**Goals.**
- 🎯 **This week:** endorser confirmed (or bioRxiv fallback ready); Paper 2 reviewed twice by P10 + one human; one-screen site live; 5 applications sent; scanpy tutorial run end to end.
- 🏁 **Next milestone:** Gate 20.
- 🌟 **Final:** a public identity that matches the work.

**The plan (v5, unchanged).**
- **Goal:** endorser confirmed; draft reviewed; one-screen site.
- **Tue (B19):** scanpy PBMC tutorial.
- **Thu (M19):** effect sizes and intervals across both papers.
- **Build:** [ ] Mon: close the endorsement (Part 9.5); if none yet, bioRxiv fallback prepared. [ ] Wed: Prompt P10 twice + human read (same reviewer as Paper 1, or the endorser). [ ] Fri: one-screen personal site: two preprints, one instrument, one dataset, the gauge level, contact.
- **Sun:** L-Sun. L-Wed. Employee track: 5 applications to scientific-data or bioprocess-AI companies with the two preprints as the cover letter.
- **Only one thing:** the endorser's yes.

**Learn it.**
- ▶ **Core**
  - Tue · scanpy "Preprocessing and clustering 3k PBMCs" tutorial (scanpy.readthedocs.io, tutorials section) (75 min, run it).
  - Thu · *Modern Statistics for Modern Biology* sections on reporting and ISLP ch. 5.2 (bootstrap) re-read (40 min); apply bootstrap intervals across both papers.
  - StatQuest "PCA main ideas", "t-SNE, Clearly Explained" and "UMAP Dimension Reduction, Main Ideas!!!" (YouTube, 30 min) — the three plots every single-cell talk shows.
- ＋ **Deeper**
  - *Single-cell best practices* (Heumos et al., theislab) — sc-best-practices.org, introduction and quality-control chapters.
  - MIT 6.874 single-cell lecture (Kellis, YouTube).
  - Cumming (2014), "The new statistics: why and how", *Psychological Science* 25(1):7-29 — the case for estimation and intervals over p-values.
  - Chip Huyen, *ML Interviews Book* — huyenchip.com/ml-interviews-book (free) — chapters on the interview process, for the applications you send this week.

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| scverse/scanpy | ★ 2.6k | PBMC 3k tutorial notebook (in the docs; `scverse/scanpy-tutorials` is archived, ★ 242) in Colab |
| theislab/single-cell-best-practices | ★ 1.3k | Run the quality-control notebook on the same data; compare cell counts after filtering |
| chiphuyen/ml-interviews-book | ★ 4.8k | Read the "career" chapters; write 3 STAR stories from your preprints |
| scipy/scipy | ★ 15.0k | `scipy.stats.bootstrap` on your key metrics; replace every naked number in both papers |

**NotebookLM kit · `LABLOOP-W19 · Recruiter portfolio and review`**
- **Load (10):** (1) this chapter; (2) scanpy PBMC tutorial page; (3) sc-best-practices intro; (4) StatQuest t-SNE/UMAP; (5) Cumming 2014; (6) Paper 1 bioRxiv PDF; (7) Paper 2 draft v2; (8) human review notes; (9) site text; (10) 5 job descriptions you applied to (pasted).
- **Focus line:** `Focus: reading single-cell analyses (QC, PCA, clustering, UMAP), reporting every number with an interval, and presenting a scientific portfolio to scientific-data companies.`
- **Custom question:** `For each of the five job descriptions, map requirements to evidence in my two preprints and repos. Where is the biggest gap and what artifact would close it?`

**Mentor check.** *"Someone who never met me can open one page and see two preprints, one instrument, one dataset and how to reach me."*

---

### Week 20 · GATE 20: Paper 2 on arXiv, endorsed (L4-endorsed)

**Act IV (finale) · Threads: B, M, BL, R · GATE 20**

**Welcome.** Halfway. Stop and look back down the mountain. Twenty weeks ago you had a cockpit and a headline. Today you submit your second preprint with a scientist's endorsement, you propose a talk, and you plan Half 2. Two mock sessions close the act: on Tuesday a principal investigator grills you in a simulated lab meeting; on Thursday a staff ML scientist interviews you. They are rehearsals, like a pilot's check ride before the licence is upgraded. The public retrospective matters too. Program managers know the value of a good lessons-learned session; the difference now is that you publish it, which is how scientists build trust in each other: by showing the process, not just the result.

**Where this week sits.**
- ← **From:** all of Half 1.
- → **Feeds:** all of Half 2: Paper 2 is your calling card for the partner (wk 21), the Loop Audit offer (wk 27), established companies (Part 9.4), and the journal submission (wk 36); the talk proposal becomes week 38.
- ⟳ **Returns in:** week 40's final retro compares against this one.

**Goals.**
- 🎯 **This week:** arXiv submission (endorsed) or bioRxiv with dated endorsement conversation; talk proposal sent; public retro; Half 2 lines in `PROGRESS.md`.
- 🏁 **This gate (20):** L4 endorsed.
- 🌟 **Final:** the second half starts with two preprints, three instruments in motion and a network.

**The plan (v5, unchanged).**
- **Goal:** submitted, endorsed; talk proposed; Half 2 planned.
- **Tue (B20):** mock lab meeting (Prompt P3).
- **Thu (M20):** mock ML interview (Prompt P3).
- **Build:** [ ] Mon: arXiv submission (cs.LG, cross-list q-bio.QM) with endorsement; bioRxiv as well if the endorser is late. [ ] Wed: talk proposal to a biotech event in Buenos Aires or Rosario, or an online LATAM one: "Two thousand fermentation papers and a hundred open batches: what the data says". [ ] Fri: public retro (what the gauge says, what slipped); Half 2 `PROGRESS.md` lines written.
- **Sun:** L-Sun retro. **X2**. Gauge: **L4 (endorsed)**.
- **Gate 20 passes when:** arXiv ID exists, or bioRxiv ID exists with an endorsement conversation dated.
- **Only one thing:** the submission.

**Learn it.**
- ▶ **Core**
  - Tue · Re-read all biology notes; P3 with the "lab meeting PI" persona in `LABLOOP-W20` (75 min).
  - Thu · P3 with the "staff ML scientist at a bioprocess company" persona; record a 10-minute explanation of Paper 2 (75 min).
  - Patrick Winston, "How to Speak" (MIT OCW / YouTube, ~1 h) — watch Saturday; it prepares both the mock sessions and the talk proposal.
- ＋ **Deeper**
  - Richard Hamming, *You and Your Research* (re-read from week 0; you will hear it differently now).
  - Stanford CS197 lecture on research talks and posters (verify title).
  - Coursera *Introduction to Public Speaking* (Univ. of Washington), week 1.

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| marimo-team/marimo | ★ 22.7k | Turn the Paper 2 learning curve into a reactive notebook with a slider for number of batches; share it as the "try it yourself" link in the arXiv abstract |
| quarto-dev/quarto-cli | ★ 6.0k | Build the arXiv-ready LaTeX source from Quarto; check it compiles in arXiv's TeX Live |
| your repos | — | Tag `paper-2-v1.0` and mint a Zenodo DOI for the exact code used |

**NotebookLM kit · `LABLOOP-W20 · Halfway gate`** (gate week: add Video Overview + Slide Deck; then create `LABLOOP-ACT-IV`)
- **Load (12):** (1) this chapter; (2) Paper 1 PDF; (3) Paper 2 PDF; (4) Winston "How to Speak"; (5) your recorded explanation transcript (P18); (6) `PROGRESS.md` snapshot; (7) `LOG.md` summary of 20 weeks; (8) talk proposal; (9) Half 2 plan (Part 7 of this document); (10-12) `LABLOOP-ACT-I/II/III` Briefing docs.
- **Focus line:** `Focus: what I have built and learned in 20 weeks, the two papers explained for a lab meeting and for an ML interview, and what Half 2 must achieve.`
- **Slide Deck prompt:** `Talk proposal deck, 6 slides: "Two thousand fermentation papers and a hundred open batches: what the data says" — hook, two findings with numbers, one method, one limitation, why it matters for LATAM biotech.`
- **Custom question:** `Compare my week 0 goals with my week 20 evidence. What did I underestimate, what did I overestimate, and what must Half 2 protect at all costs?`

**Mentor check.** *"I am an author of two preprints, one endorsed. I can present both without notes to a biologist and to an ML scientist."*

---

## PART 7 · Weeks 21-40 as chapters (depth, co-author, money)

Same weekly shape as Part 6. Study tracks become lighter because the build blocks now carry Paper 3, the instruments and the money thread.

---

# ACT V · The protein detective finds a partner (weeks 21-24)

> **Act V story.** A detective can read fingerprints faster than anyone, but a fingerprint match is not a conviction. The case closes only when a forensic lab runs the test. Protein language models are your fingerprint reader: they turn each enzyme sequence into an "embedding" that carries hints about how it behaves, including how much heat it tolerates. Your job in Act V is to become a careful detective (honest baselines, no leakage between protein families, intervals) and, above all, to find the forensic lab: a founder-scientist who will express and test your top candidates. When they agree, their name goes next to yours. That is L4 co-authored, the moment you stop being a vendor.
>
> **Format note for weeks 21-40.** v5 Part 7 gave these weeks as one table row each. Each chapter below expands that row into the v5 weekly format (Goal · Study · Build · Output · Only one thing) **without changing the gate or the output**. The original row is quoted first.

---

### Week 21 · Paper 3 scoped with a partner

**Act V · Threads: B, EL, R · Next gate: 24**

**Welcome.** Picking a scientific problem is like picking a hiking trail with a friend who knows the mountains. You bring the map-making skills (embeddings, baselines, evaluation); they bring the terrain (the enzyme family they care about, the organisms in their freezer, the assays they can run). The best trail sits where your feasibility and their interest overlap, which is exactly Uri Alon's chart from week 0. This week you set up `enzymelens`, cache ESM-2 embeddings for the full Meltome split overnight on your 3050 (your GPU finally earns its keep), fit three simple heads (ridge, gradient boosting, small MLP), and book the partner call with your two preprints as the calling card. Deep-read FLIP and Meltome with the partner's enzyme classes in mind; you want to walk into the call already speaking their dialect.

**Where this week sits.**
- ← **From:** week 16-17 (ESM-2, Meltome baseline), week 18 (extremophile founder digests), Part 9.3 ladder (rungs 1-4 with those founders).
- → **Feeds:** week 22 (honest baselines), week 23 (partner-specific candidates), week 24 (co-author agreement), weeks 30-32 (Paper 3).
- ⟳ **Returns in:** Paper 3 Introduction and Methods.

**Goals.**
- 🎯 **This week:** `enzymelens` repo; embeddings cached; baseline table for three heads; partner call booked.
- 🏁 **Next milestone:** Gate 24 (written co-author agreement; pre-registered validation plan).
- 🌟 **Final:** Paper 3 with a founder co-author and wet-lab validation (wk 32).

**The plan (from v5 Part 7, expanded; gate and output unchanged).**
> *v5 row:* 21 | Paper 3 scoped with a partner | Meltome and FLIP deep read; extremophile enzyme families of the 3 candidate partners (Puna Bio, Antarka, APEXzymes) | `enzymelens` repo; ESM-2 embeddings for the full Meltome split cached on disk (overnight on the 3050); ridge/GBM/MLP heads; baseline table | Partner call booked with the FermLit + batchlens preprints as the calling card
- **Goal:** Paper 3 has a candidate partner, a dataset on disk and a baseline table.
- **Tue (study):** FLIP thermostability task and Meltome atlas, deep read; for each of the 3 candidate partners, list the enzyme families (EC classes) and organisms they publish on.
- **Build:** [ ] Mon: `enzymelens` skeleton (same README template as Part 5.2); download Meltome splits; start overnight embedding job with checkpointing. Done when the job resumes after an interruption. [ ] Wed: ridge, LightGBM and a 2-layer MLP on cached embeddings; Spearman ρ with bootstrap intervals. Done when the baseline table regenerates with `make baselines`. [ ] Fri: partner email (P13, co-authorship variant) sent to the most engaged founder, with both preprints and one figure relevant to their enzyme class. Done when the call is booked or a follow-up date is logged.
- **Sun:** L-Sun "What a protein language model sees in 30,000 melting curves" with the baseline table.
- **Only one thing:** the partner call booked.

**Learn it.**
- ▶ **Core**
  - Tue · Jarzab et al. 2020 (Meltome), methods and results; Dallago et al. 2021 (FLIP), thermostability task section (75 min).
  - Uri Alon (2009), "How to choose a good scientific problem" (third reading; this time fill his feasibility/interest chart for Paper 3) (15 min).
  - iBiology, "Uri Alon: How to Choose a Good Scientific Problem" (verify; talk version).
- ＋ **Deeper**
  - Pudžiuvelytė et al. (2024), "TemStaPro: protein thermostability prediction using sequence representations from protein language models", *Bioinformatics* 40(4):btae157 — the closest published tool to what you are building.
  - Yang, Wu & Arnold (2019) review (carry over).
  - Valence Labs (YouTube) reading-group talks on protein language models.
  - Frances Arnold on iBiology (directed evolution talks; verify).

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| ievapudz/TemStaPro | ★ 87 (niche; the published tool's official repo) | Run it on 50 Meltome sequences; compare its predictions with your ridge head. This is your most relevant external baseline |
| J-SNACKKB/FLIP | ★ 141 (niche, official) | Use their exact splits so your numbers are comparable with the paper |
| huggingface/transformers | ★ 165.3k | `EsmModel` with `torch_dtype=torch.float16`; batch by sequence length to fit 6 GB |
| agemagician/ProtTrans | ★ 1.3k | Read the README; note ProtT5 as an alternative embedding for the ablation you might run in week 22 |
| biopython/biopython | ★ 5.2k | `Bio.SeqIO` to read and length-filter FASTA; drop sequences > 1,022 residues or chunk them |

**NotebookLM kit · `LABLOOP-W21 · Protein study and partner scope`**
- **Load (11):** (1) this chapter; (2) Meltome atlas; (3) FLIP preprint; (4) TemStaPro paper; (5) Yang/Wu/Arnold 2019; (6) Alon 2009; (7) 3 partner-founder paper digests (P11); (8) partner company public pages (as websites); (9) baseline table; (10) your partner email; (11) Paper 1 and Paper 2 abstracts.
- **Focus line:** `Focus: thermostability as a measurable protein property, how protein language model embeddings are used with simple prediction heads, and how to scope a joint study with a wet-lab partner.`
- **Custom question:** `Prepare me for the partner call: five questions to ask about their enzymes and assays, three things I must not promise, and one pilot scope that fits 8 weeks.`

**Mentor check.** *"I can explain what my baseline predicts, how well, and why the partner's lab is the only way to know if it matters."*

---

### Week 22 · Baselines beaten honestly

**Act V · Threads: M, EL, R · Next gate: 24**

**Welcome.** There is an old trick in sales: compare your product with a weak competitor and you always win. Science has the same trick, and reviewers hunt for it. The honest move is to compare against the strongest published baseline, on the exact same splits, and to split proteins **by family** so the model cannot cheat by recognising a close cousin it saw in training (like a student who memorised last year's exam because this year's is almost identical). This week you do exactly that, add conformal intervals, and compare with the FLIP published numbers. You also start one lecture on causal inference, because in Paper 3 the difference between "correlated with thermostability" and "causes thermostability" is the difference between a candidate list and a claim.

**Where this week sits.**
- ← **From:** week 21 baselines, week 16 conformal prediction, week 4 grouped CV.
- → **Feeds:** week 23 (candidate ranking with uncertainty), week 24 (pre-registration defines success against these baselines), week 30 (validation analysis), week 31 (Paper 3 Methods).
- ⟳ **Returns in:** Paper 3 Table 1.

**Goals.**
- 🎯 **This week:** family-split CV; conformal intervals; comparison to FLIP published baselines; L-Sun with the table.
- 🏁 **Next milestone:** Gate 24.
- 🌟 **Final:** a Paper 3 no reviewer can call "leaky".

**The plan (from v5 Part 7, expanded; gate and output unchanged).**
> *v5 row:* 22 | Baselines beaten honestly | Brady Neal causal inference lecture 1 (for confounders in Paper 3's analysis) | Cross-validation by protein family (no leakage); conformal intervals; compare to FLIP published baselines | L-Sun with the table
- **Goal:** a baseline table a sceptical reviewer accepts.
- **Tue (study):** Brady Neal, *Introduction to Causal Inference*, lecture 1 + book ch. 1 (bradyneal.com/causal-inference-course).
- **Build:** [ ] Mon: cluster sequences by identity (e.g., MMseqs2 at 30-50% identity) and build grouped folds. Done when no cluster appears in two folds. [ ] Wed: re-run the three heads on grouped folds; conformal intervals; coverage. Done when the table shows random-split vs family-split side by side. [ ] Fri: FLIP published numbers in the same table, same metric; write the three-line honest interpretation. Done when `make baselines` regenerates it.
- **Sun:** L-Sun "Why my protein model got worse when I stopped letting it cheat" with the table.
- **Only one thing:** the family-split table.

**Learn it.**
- ▶ **Core**
  - Tue · Brady Neal, *Introduction to Causal Inference*, lecture 1 (YouTube) and book ch. 1 (free PDF on the course site) (75 min).
  - StatQuest "Spearman's Rank Correlation" if needed (search; verify) and "Bootstrapping Main Ideas" (carry over).
- ＋ **Deeper**
  - Cunningham, *Causal Inference: The Mixtape* — mixtape.scunning.com, ch. 3 (directed acyclic graphs).
  - Huntington-Klein, *The Effect* — theeffectbook.net, ch. 6-7 (causal diagrams), friendlier tone.
  - Coursera *A Crash Course in Causality* (Penn), week 1.
  - Steinegger & Söding (2017), "MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets", *Nature Biotechnology* 35:1026-1028.

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| py-why/dowhy | ★ 8.3k | Run the "getting started" notebook; draw a DAG for "organism growth temperature → sequence composition → Tm" to see where confounding lives |
| scikit-learn/scikit-learn | ★ 67.2k | `GroupKFold` / `StratifiedGroupKFold` with sequence-cluster IDs |
| scikit-learn-contrib/MAPIE | ★ 1.6k | Conformal intervals on the grouped folds; report coverage per fold |
| soedinglab/MMseqs2 | ★ 2.1k | `mmseqs easy-cluster` on Meltome FASTA at 30% identity; save cluster TSV |

**NotebookLM kit · `LABLOOP-W22 · Protein validation and baselines`**
- **Load (9):** (1) this chapter; (2) Brady Neal lecture 1 + ch. 1; (3) The Effect ch. 6; (4) FLIP preprint (baselines table); (5) MMseqs2 paper abstract; (6) Kapoor & Narayanan leakage paper; (7) random vs family split table; (8) coverage results; (9) week 21 notebook Briefing doc.
- **Focus line:** `Focus: data leakage through protein homology, why family splits are the honest test, conformal intervals on protein predictions, and confounding in thermostability data.`
- **Custom question:** `Draw (in words) the causal diagram linking source organism, growth temperature, sequence features and measured Tm. Which confounder could make my model look better than it is?`

**Mentor check.** *"I can explain why my score dropped with family splits, and why the lower number is the one I publish."*

---

### Week 23 · Extremophile focus

**Act V · Threads: B, EL, R · Next gate: 24**

**Welcome.** Until now you trained a detective on a general catalogue of fingerprints. This week you go to the partner's neighbourhood. Extremophiles live where most life would die: boiling springs, salt flats, Antarctic ice. Their enzymes are tough, which is exactly what industry wants (a detergent enzyme that works in hot water, a process enzyme that survives a 70 °C reactor). You pull sequences from UniProt for the partner's enzyme class, predict thermostability, and produce a ranked candidate list with uncertainty. Think of it as a shortlist from a recruiter: ten candidates, each with a score, a confidence and a one-line reason, handed to the hiring manager (the partner) who runs the real interview (the assay). You share it under a simple collaboration note so everyone knows who owns what.

**Where this week sits.**
- ← **From:** weeks 21-22 (honest model), week 18 and 21 (partner digests and call).
- → **Feeds:** week 24 (validation plan picks 5-10 of these), week 28-30 (wet-lab results), week 31 (Paper 3 Results).
- ⟳ **Returns in:** the partner's lab bench, which is the only place Paper 3 can become a discovery.

**Goals.**
- 🎯 **This week:** partner-class sequences from UniProt; predictions with intervals; ranked candidate list; collaboration note sent.
- 🏁 **Next milestone:** Gate 24.
- 🌟 **Final:** candidates tested in a real lab, reported in Paper 3.

**The plan (from v5 Part 7, expanded; gate and output unchanged).**
> *v5 row:* 23 | Extremophile focus | Partner's organisms and enzyme classes | Filter and enrich: sequences from UniProt for the partner's enzyme class; predict thermostability; ranked candidate list with uncertainty | Candidate list shared with the partner under a simple collaboration note
- **Goal:** a shortlist the partner can act on.
- **Tue (study):** the partner's organisms and enzyme classes: two review papers or book chapters the partner recommends (ask in the call), plus the partner's own last two papers.
- **Build:** [ ] Mon: UniProt REST query for the enzyme class (EC number) filtered by taxonomy (thermophiles/psychrophiles as relevant); FASTA + metadata. Done when the query is saved in the repo and reproducible. [ ] Wed: embeddings + predictions + conformal intervals; flag sequences outside the training distribution (nearest-neighbour identity to Meltome). Done when every candidate has score, interval and novelty. [ ] Fri: ranked list (top 20) with a one-line rationale each; collaboration note (one paragraph: data ownership, authorship intent, no confidential data in public repo). Done when sent.
- **Sun:** L-Sun (no partner names without permission): "From 30,000 proteins to 20 candidates: what a shortlist with uncertainty looks like".
- **Only one thing:** the ranked list sent.

**Learn it.**
- ▶ **Core**
  - Tue · Partner-recommended reading (ask for it); if none yet, Rothschild & Mancinelli (2001) "Life in extreme environments" (carry over) and the UniProt "Help" pages on EC numbers and taxonomy filters (75 min).
  - Khan Academy / Crash Course refresh on protein folding and denaturation ("why heat unfolds proteins") (15 min).
- ＋ **Deeper**
  - UniProt REST API documentation (rest.uniprot.org) and the BRENDA enzyme database (brenda-enzymes.org) for reported temperature optima.
  - iBiology talks on extremophiles and enzymes (search; verify).
  - Crash Course Biology #35 "Old & Odd" (carry over).

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| biopython/biopython | ★ 5.2k | Parse UniProt FASTA, compute length and composition features as a sanity baseline next to embeddings |
| sokrypton/ColabFold | ★ 2.9k | Colab: predict the structure of your top 2 candidates; look at them in 3D. Not used as a feature, only to build intuition and a figure |
| google-deepmind/alphafold | ★ 14.9k | Read the README on the AlphaFold Protein Structure Database; check if your candidates already have predicted structures there (no GPU needed) |
| deepchem/deepchem | ★ 7.0k | Browse "The Basic Tools of the Deep Life Sciences" tutorial list; one notebook on protein sequences, only if curious (Saturday) |

**NotebookLM kit · `LABLOOP-W23 · Extremophile candidate ranking`**
- **Load (10):** (1) this chapter; (2) partner-recommended papers; (3) partner's last two papers; (4) Rothschild & Mancinelli 2001; (5) UniProt help pages (EC, taxonomy); (6) TemStaPro paper; (7) your ranked list with intervals; (8) novelty analysis; (9) collaboration note; (10) week 22 Briefing doc.
- **Focus line:** `Focus: extremophile enzymes and why thermostability matters industrially, turning model predictions with uncertainty into a ranked shortlist, and how to share it responsibly with a partner.`
- **Studio extra (Infographic):** `How a candidate list is made: 30,000 Meltome proteins → model → partner enzyme class from UniProt → predictions with intervals → novelty filter → top 20 → lab test.`
- **Custom question:** `For my top 5 candidates, which are "safe bets" and which are "long shots", based on intervals and distance from training data? How should the partner choose a mix?`

**Mentor check.** *"I can hand a scientist 20 candidates and explain, for each one, why it is there and how sure I am."*

---

### Week 24 · GATE 24: co-authorship agreed

**Act V (finale) · Threads: B, EL, R · GATE 24**

**Welcome.** In a well-run joint venture, the most important document is signed before any money moves: who contributes what, who owns what, what counts as success. Science has the same document, smaller and friendlier: a paragraph on authorship and a pre-registered validation plan. Pre-registration is your old friend "acceptance criteria before the sprint", except you publish it in the repo **before** the lab results arrive, so nobody (including you) can move the goalposts afterwards. This week you also sit in the student's chair: the partner explains how wet-lab validation works (expressing an enzyme, purifying it, measuring activity at different temperatures), and you take notes like a new analyst in a plant visit. That humility is exactly why they will want your name next to theirs.

**Where this week sits.**
- ← **From:** weeks 21-23 (model, shortlist, partner relationship); week 6 (you already know how to write a protocol before data).
- → **Feeds:** weeks 28-30 (validation results analysed against this plan), weeks 31-32 (Paper 3), Part 9 ("co-author with [lab]" in the headline).
- ⟳ **Returns in:** L4 co-authored at week 32; the credibility behind every Loop Audit.

**Goals.**
- 🎯 **This week:** validation plan with 5-10 candidates and a success definition, committed to the repo with a date; written co-author agreement (one paragraph, emailed).
- 🏁 **This gate (24):** co-author agreed; plan pre-registered.
- 🌟 **Final:** Paper 3 with a founder co-author; the peer status of L4.

**The plan (from v5 Part 7, expanded; gate and output unchanged).**
> *v5 row:* 24 | Co-authorship agreed | How wet-lab validation works (partner explains; you take notes) | Validation plan: which 5-10 candidates the partner can express and test in their lab; what "success" is, pre-registered in the repo | Written co-author agreement (one paragraph, emailed)
- **Goal:** a signed intention and a frozen plan.
- **Tue (study):** a session with the partner on their validation workflow (expression host, purification, activity assay, temperature series); your notes in `enzymelens/validation/PROTOCOL-NOTES.md`.
- **Build:** [ ] Mon: draft `PREREGISTRATION.md`: hypotheses, candidates, controls, primary metric (e.g., residual activity after X min at Y °C, or measured Tm), success threshold, analysis plan for small n. Done when the partner has commented. [ ] Wed: incorporate comments; commit with date and tag `prereg-v1`; optional OSF registration. Done when the tag exists. [ ] Fri: authorship paragraph emailed (who does what, author order intention, data sharing, what happens if results are negative: still published). Done when the partner replies yes.
- **Sun:** L-Sun "I pre-registered my first wet-lab validation. Here is why an engineer should love that" (no confidential details). Gauge: L4 (path).
- **Gate 24 passes when:** written co-author agreement exists; validation plan is pre-registered in the repo.
- **Only one thing:** the partner's written yes.

**Learn it.**
- ▶ **Core**
  - Tue · The partner session (60 min) + Nosek et al. (2018), "The preregistration revolution", *PNAS* 115(11):2600-2606 (30 min).
  - LabXchange simulations on protein expression/purification or enzyme assays (search the library; verify availability) so the partner's words land on a picture.
- ＋ **Deeper**
  - McNutt et al. (2018), "Transparency in authors' contributions and responsibilities to promote integrity in scientific publication", *PNAS* 115(11):2557-2560 (CRediT roles).
  - CRediT taxonomy (credit.niso.org) — the standard list of contributor roles; use it in the authorship paragraph.
  - Center for Open Science / OSF preregistration templates (osf.io).
  - Coursera *Research Data Management and Sharing*, module on data sharing agreements.

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| lmfit/lmfit-py | ★ 1.2k | Build the analysis you pre-registered on simulated numbers: fit a thermal inactivation curve, get intervals; the code is frozen before real data arrives |
| scipy/scipy | ★ 15.0k | Power check by simulation: with n = 3 replicates per candidate, how big a difference can you detect? Write the answer into the plan |
| the-turing-way/the-turing-way | ★ 2.2k | Read the chapter on "Collaboration" and "Credit"; adapt its checklist for the authorship paragraph |

**NotebookLM kit · `LABLOOP-W24 · Co-author and validation gate`** (gate week: add Video Overview + Slide Deck; then create `LABLOOP-ACT-V`)
- **Load (10):** (1) this chapter; (2) Nosek 2018; (3) McNutt 2018; (4) CRediT taxonomy page; (5) OSF preregistration template page; (6) your protocol notes; (7) `PREREGISTRATION.md`; (8) authorship paragraph; (9) power simulation result; (10) weeks 21-23 Briefing docs.
- **Focus line:** `Focus: how wet-lab validation of enzyme candidates works, pre-registration as acceptance criteria published before results, and fair authorship with CRediT roles.`
- **Slide Deck prompt:** `Five slides for the partner's team meeting: the question, the candidates, what we will measure, what counts as success, who does what.`
- **Custom question:** `Stress-test my pre-registration: list every way results could be analysed "flexibly" after the fact that my plan does not yet forbid.`

**Mentor check.** *"A scientist has agreed in writing to co-author with me, and the definition of success is frozen in my repo before any data exists."*

---


# ACT VI · The first client (weeks 25-32)

> **Act VI story.** The cartographer opens a small office on the town square. On the wall: two published maps (your preprints) and one being drawn with a local guide (Paper 3). On the desk: two survey instruments (batchlens, and this act's new one, trialkit). On the door: a price list with one item, the Loop Audit, fixed price, two weeks. Townspeople do not buy maps; they buy fewer wasted journeys. Act VI is where "I measure loops" becomes "I am paid to measure your loop". While the office opens, the partner's lab is running your candidates; results arrive mid-act, and Paper 3 is written with them.

---

### Week 25 · Instrument 2 trialkit (field module)

**Act VI · Threads: M, $, R · Next gate: 28**

**Welcome.** Picture an agronomist with 40 spreadsheets from three seasons of field trials, each formatted by a different intern, and an investor asking "does your biological product actually increase yield, and by how much?". The honest answer requires a design (which plots got which treatment, in which block), a statistical model that respects that design, and an interval. Most startups answer with an average and a hopeful chart. Trialkit's field module is the translator: an LLM maps messy spreadsheets into one trial schema (the same trick as FermLit, now on tables), and a mixed model produces treatment effects with intervals. The mixed model is easier than it sounds: fields differ (some blocks are sunnier), and the model simply says "compare treatments *within* each block, then combine", the way you compare sales teams within regions before comparing regions. This is v4's weeks 9-10, now standing on two preprints.

**Where this week sits.**
- ← **From:** week 7 (LLM mapping measured against a gold set), week 14 (multiple testing), v4 Part 5.2 (trialkit design).
- → **Feeds:** week 26 (plate module), week 27 (the offer shows two instruments), week 29 (audit prototype on client data), archetype B companies in v4 Appendix A.
- ⟳ **Returns in:** Loop Audits for field-trial companies; the Medium deep dive planned in v4.

**Goals.**
- 🎯 **This week:** `trialkit/ingest` (schema mapping with measured accuracy) and `trialkit report` (design detection, mixed model, intervals); `trialkit v0.1` + Z5.
- 🏁 **Next milestone:** Gate 28 (first invoice).
- 🌟 **Final:** a second instrument that widens who can buy a Loop Audit.

**The plan (from v5 Part 7, expanded; gate and output unchanged).**
> *v5 row:* 25 | Instrument 2 trialkit (field module) | NIST ch. 5 (DoE); statsmodels MixedLM | `trialkit/ingest` and `report` (v4 weeks 9-10 content) | `trialkit v0.1`, Z5
- **Goal:** from 10 heterogeneous published trial spreadsheets to one defensible report.
- **Tue (study):** NIST/SEMATECH e-Handbook ch. 5.1-5.3 (DoE: randomised block designs, factorials); statsmodels MixedLM documentation.
- **Build:** [ ] Mon: trial schema (plot, block, treatment, replicate, date, variable, unit); LLM mapping of 10 `agridat` trials (via the Rdatasets mirror) with Claude vs local Qwen; mapping accuracy against 10 hand-mapped files. Done when the accuracy table prints. [ ] Wed: design detection (CRD, RCBD, split-plot) and MixedLM with treatment effects and intervals; bootstrap check. Done when a known published trial's effect is reproduced within its reported interval. [ ] Fri: `trialkit report` (HTML, Spanish and English); release v0.1 + Zenodo DOI. Done when the DOI badge renders.
- **Sun:** L-Sun "10 field trials, 10 spreadsheet formats, one schema, one defensible number". Z5.
- **Only one thing:** one published trial reproduced with an interval.

**Learn it.**
- ▶ **Core**
  - Tue · NIST e-Handbook ch. 5.1-5.3 — itl.nist.gov/div898/handbook (45 min); Penn State STAT 503 *Design of Experiments* online notes, lesson on randomised complete block designs — online.stat.psu.edu/stat503 (30 min).
  - Piepho, Büchse & Emrich (2003), "A hitchhiker's guide to mixed models for randomized experiments", *Journal of Agronomy and Crop Science* 189:310-322 (skim sections 1-3).
- ＋ **Deeper**
  - Michael Clark, *Mixed Models with R* — m-clark.github.io/mixed-models-with-R (concepts transfer to Python).
  - Bambi documentation (bambinos.github.io/bambi), Bayesian mixed models in one formula.
  - StatQuest on ANOVA and "Design matrices" (search in channel).
  - Coursera *Practical Time Series Analysis* — only if a client's trials are longitudinal.

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| statsmodels/statsmodels | ★ 11.6k | MixedLM example from the docs (random intercept per block); then fit one `agridat` trial |
| bambinos/bambi | ★ 1.3k | Same trial as a Bayesian mixed model; compare intervals with MixedLM |
| vincentarelbundock/Rdatasets | ★ 416 (widely used data mirror) | Pull 10 `agridat` CSVs by URL; this is your public data source |
| 567-labs/instructor | ★ 13.9k | Spreadsheet header → schema mapping with a Pydantic model, as in FermLit |
| m-clark/mixed-models-with-R | ★ 154 (niche, excellent free book) | Read the "random intercepts" chapter for intuition |

**NotebookLM kit · `LABLOOP-W25 · trialkit field module`**
- **Load (10):** (1) this chapter; (2) v4 Part 5.2 text; (3) NIST ch. 5 pages; (4) PSU STAT 503 RCBD lesson; (5) Piepho 2003; (6) statsmodels MixedLM docs; (7) Bambi docs page; (8) trial schema; (9) mapping accuracy table; (10) reproduced trial result.
- **Focus line:** `Focus: randomised block designs and mixed models explained as comparing within regions before comparing regions, and LLM schema mapping for messy field-trial spreadsheets.`
- **Custom question:** `A grower asks: "is a 7% yield increase real?" Using my trialkit output, write the three-sentence answer a statistician would sign.`

**Mentor check.** *"I can take a messy field trial and say the treatment effect with an interval that respects the design."*

---

### Week 26 · Instrument 2 (plate module)

**Act VI · Threads: B, M, $, R · Next gate: 28**

**Welcome.** Walk into a microbiology lab and you will see plate readers: machines that measure 96 or 384 tiny wells every few minutes. Out comes a spreadsheet with thousands of numbers, and someone spends an afternoon in Excel fitting curves by eye. The plate module ends that afternoon. Growth curves (the same lag-log-stationary story from week 1, now in tiny wells), dose-response curves (how much of a compound it takes to halve growth, the famous IC50), Michaelis-Menten (week 7 again), and a quality number called Z′ that says whether the assay itself is good enough to trust. Z′ is like checking that a scale reads zero with nothing on it and reads 1 kg with a 1 kg weight, before you weigh anything that matters. This week also quietly prepares your eye for Paper 3's validation data, which will come from similar assays.

**Where this week sits.**
- ← **From:** week 1 (growth phases), week 7 (Michaelis-Menten), week 16 (intervals), week 25 (trialkit schema and report).
- → **Feeds:** week 27 (offer covers archetype C, lab assays), weeks 28-30 (partner's validation data analysis), Loop Audits for assay-heavy startups.
- ⟳ **Returns in:** Paper 3 Results (enzyme activity curves).

**Goals.**
- 🎯 **This week:** `trialkit/plate`: parse plate-reader exports, growth curves, 4PL dose-response, Michaelis-Menten, Z′ QC, with intervals.
- 🏁 **Next milestone:** Gate 28.
- 🌟 **Final:** the same statistical care applied to the partner's validation data.

**The plan (from v5 Part 7, expanded; gate and output unchanged).**
> *v5 row:* 26 | Instrument 2 (plate module) | Assay statistics (4PL, growth curves, Z′) | `trialkit/plate` (v4 week 11) | L-Sun
- **Goal:** a plate export in, a QC'd, fitted, interval-bearing report out.
- **Tue (study):** assay statistics: 4PL, logistic/Gompertz growth models, Z′-factor.
- **Build:** [ ] Mon: parser for 96/384 layouts (plate map + readings) into the trial schema. Done when two different vendor export formats load. [ ] Wed: growth-curve fits (logistic, Gompertz; growth rate, lag, carrying capacity) and 4PL dose-response with bootstrap intervals. Done when a published dataset's parameters are reproduced. [ ] Fri: Z′ per plate and Michaelis-Menten module; plate report. Done when a bad plate is flagged automatically.
- **Sun:** L-Sun "The Z′ factor: the one number that says whether your assay deserves your conclusions".
- **Only one thing:** growth-curve fits with intervals.

**Learn it.**
- ▶ **Core**
  - Tue · Zhang, Chung & Oldenburg (1999), "A simple statistical parameter for use in evaluation and validation of high throughput screening assays", *Journal of Biomolecular Screening* 4(2):67-73 (20 min); Zwietering et al. (1990), "Modeling of the bacterial growth curve", *Applied and Environmental Microbiology* 56(6):1875-1881 (30 min); Ritz et al. (2015), "Dose-response analysis using R", *PLoS ONE* 10(12):e0146021, sections on the 4-parameter log-logistic model (25 min).
- ＋ **Deeper**
  - Sprouffske & Wagner (2016), "Growthcurver: an R package for obtaining interpretable metrics from microbial growth curves", *BMC Bioinformatics* 17:172.
  - Midani, Collins & Britton (2021), "AMiGA: software for automated analysis of microbial growth assays", *mSystems* 6:e00508-21 — an open-source Python tool that fits growth curves without assuming their shape (`pip install amiga`).
  - NCATS *Assay Guidance Manual* (ncbi.nlm.nih.gov/books/NBK53196), chapters on assay validation and HTS statistics — the industry reference, free.
  - Khan Academy / LabXchange on spectrophotometry and optical density.

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| lmfit/lmfit-py | ★ 1.2k | Fit logistic and Gompertz to a public growth curve with parameter intervals |
| scipy/scipy | ★ 15.0k | `curve_fit` for 4PL; compare with lmfit output |
| DoseResponse/drc | ★ 29 (niche; the reference R package behind Ritz et al. 2015) | Read the README and one vignette to check your Python 4PL reproduces `drc` on the same data |
| Pioreactor/pioreactor | ★ 149 (niche) | Look at how it estimates growth rate from optical density online; compare with your offline fit |

**NotebookLM kit · `LABLOOP-W26 · trialkit plate module`**
- **Load (10):** (1) this chapter; (2) Zhang 1999; (3) Zwietering 1990; (4) Ritz 2015; (5) Growthcurver paper; (6) NCATS Assay Guidance Manual validation chapter; (7) v4 week 11 text; (8) your fit results table; (9) Z′ per plate output; (10) week 1 growth-curve notes.
- **Focus line:** `Focus: plate-reader assays, growth-curve models, 4PL dose-response and IC50, Z′ as assay quality, and why these statistics protect a startup's efficacy claims.`
- **Custom question:** `A founder shows a plate with Z′ = 0.3 and a beautiful IC50. What should I say, kindly and precisely?`

**Mentor check.** *"I can look at a plate report and say whether the assay is trustworthy before I talk about the result."*

---

### Week 27 · Loop Audit offer, built on papers

**Act VI · Threads: $, R · Next gate: 28**

**Welcome.** Here is the moment your 25 years become an unfair advantage. A scientist-founder can describe their experiments perfectly and still not know how many days each loop wastes. You can. The Loop Audit is a fixed-price, two-week diagnosis: map the experimental loop, inventory the data, measure cycle time and data readiness, run one instrument on their sample, and hand back three automation candidates with estimated days saved plus one working prototype. The pitch is not "AI". The pitch is the one your GridX contact taught you: startups burn to grow, so "one fewer week per cycle × N cycles a year = months of runway you did not have to raise". This week you rewrite the one-pager with your preprints as proof (evidence, not adjectives), send it to ten people who already know your work, and aim for three calls. You are not cold-calling; you are following up on 26 weeks of rungs on the ladder.

**Where this week sits.**
- ← **From:** v4 Part 4 (offers, pricing hypotheses), Part 9.3 ladder (founders at rungs 4-6), weeks 1-26 (all proof).
- → **Feeds:** week 28 (first invoice), week 29-30 (delivery), week 32 (second audit or pilot proposal), week 39 (fund package), week 40 (decision).
- ⟳ **Returns in:** every commercial conversation for the rest of your career: the one-pager evolves, the logic stays.

**Goals.**
- 🎯 **This week:** offer one-pager rewritten with preprints as proof; sent to 10 (5 founders who engaged, 3 established companies, 2 funds); 3 calls booked.
- 🏁 **Next milestone:** Gate 28 (first invoice, any amount).
- 🌟 **Final:** a sustainable consulting practice or a fractional role (week 40 decision).

**The plan (from v5 Part 7, expanded; gate and output unchanged).**
> *v5 row:* 27 | Loop Audit offer, built on papers | v4 Part 4 offers, re-read | Offer one-pager rewritten with the two preprints as proof; sent to 10 (5 founders who engaged, 3 established companies, 2 funds) | 3 calls booked
- **Goal:** three conversations about a paid audit.
- **Tue (study):** re-read v4 Part 4 (offers and "sell weeks, not savings"), v4 Template F (one-pager) and *The Mom Test* chapters 1-3.
- **Build:** [ ] Mon: one-pager v2: problem in the client's words, what the audit delivers, two weeks, fixed price hypothesis, proof (two preprints, instruments, DOIs), what it does not do. Done when a founder friend reads it in 60 seconds and repeats it back. [ ] Wed: 10 personalised messages (Part 9.6 template adapted). Done when sent and logged in `OPPORTUNITIES.csv`. [ ] Fri: discovery-call script (10 questions, Mom-Test style: past behaviour, not opinions) and a scoping sheet. Done when both fit on one page each.
- **Sun:** L-Sun: the offer, stated plainly, with the proof links.
- **Only one thing:** the 10 messages sent.

**Learn it.**
- ▶ **Core**
  - Tue · v4 Part 4 and Template F (30 min); Rob Fitzpatrick, *The Mom Test* (book, paid, ~130 pages; ch. 1-3 = 45 min).
  - Y Combinator Startup School, "How to Talk to Users" (Eric Migicovsky, YouTube, ~20 min).
- ＋ **Deeper**
  - David Maister, Charles Green & Robert Galford, *The Trusted Advisor* (book, paid) — the trust equation.
  - Alan Weiss, *Value-Based Fees* (book, paid) — why fixed price beats hourly; matches v4's rule.
  - Stanford eCorner (ecorner.stanford.edu) — search talks on "customer discovery" and "B2B sales for technical founders".
  - Bruce Booth's *Life Sci VC* blog (lifescivc.com) — how biotech investors think about runway and milestones; vocabulary for the "months of runway" pitch.

**Run it.**

| Tool / repo | ★ | Run this (20-40 min) |
|---|---|---|
| batchlens + trialkit (yours) | — | Record a 60-second demo of each running on public data (v4 Template C) and link both in the one-pager |
| marimo-team/marimo | ★ 22.7k | A "loop calculator" notebook: inputs (cycle length, cycles/year, burn/month, days saved) → months of runway; screenshot it for the one-pager |
| anthropics/claude-cookbooks | ★ 52.7k | Borrow a summarisation recipe to turn discovery-call notes into a structured scoping sheet (never with confidential data in a cloud model without consent) |

**NotebookLM kit · `LABLOOP-W27 · Loop Audit offer and pipeline`**
- **Load (10):** (1) this chapter; (2) v4 Part 4 + Template F + Part 3.3 culture filter; (3) YC "How to Talk to Users" video; (4) your *Mom Test* notes; (5) one-pager v2; (6) 10 messages; (7) Paper 1 and 2 abstracts; (8) batchlens and trialkit READMEs (web); (9) GridX archetype notes (no names of weaknesses); (10) runway calculator screenshot text.
- **Focus line:** `Focus: selling a fixed-price diagnostic to scientist-founders, customer discovery without leading questions, and translating days saved into months of runway.`
- **Custom question:** `Role-play a sceptical CSO at a Series A fermentation startup. Push back on my Loop Audit one-pager for 8 exchanges, then tell me which objection I handled worst.`

**Mentor check.** *"I can explain the Loop Audit in 30 seconds without the word AI, and three people want to talk about it."*

---

### Week 28 · GATE 28: first Loop Audit sold

**Act VI · Threads: $, EL, R · GATE 28**

**Welcome.** You have never charged for a side project. That sentence has been true for years, and this week it stops being true. The amount does not matter; the category change does. Think of a musician who played for friends for a decade and then sells the first ticket: the song is the same, the relationship is different, and so is the musician. Scope the audit, agree the price, get half upfront. If it does not close this week, week 29 is a second selling week, as planned; the calendar bends, the gate does not. In parallel, something exciting may arrive from the partner's lab: the first wet-lab results for your candidates. Log them, do not analyse them yet: the pre-registered plan from week 24 says how, and when.

**Where this week sits.**
- ← **From:** week 27 calls; weeks 1-26 proof.
- → **Feeds:** weeks 29-30 (delivery), week 32 (case study, second proposal), week 37 (pilot), week 40 (income in the retro).
- ⟳ **Returns in:** your self-image. This gate matters as much psychologically as financially.

**Goals.**
- 🎯 **This week:** one audit scoped, agreed, half paid; first invoice issued; partner results logged untouched.
- 🏁 **This gate (28):** Money 1.
- 🌟 **Final:** a practice with retainers or a well-chosen job (week 40).

**The plan (from v5 Part 7, expanded; gate and output unchanged).**
> *v5 row:* 28 | GATE 28: first Loop Audit sold | none (selling week) | Audit scoped and half paid; Paper 3 wet-lab results start arriving | **First invoice** (any amount). If not: week 29 is a second selling week
- **Goal:** the first invoice exists.
- **Tue:** no study. Use the block for follow-ups.
- **Build:** [ ] Mon: discovery calls held; notes structured into the scoping sheet. [ ] Wed: proposal (one page: objective, deliverables, two weeks, price, 50% upfront, confidentiality, case-study permission request). Done when sent. [ ] Fri: invoice issued on acceptance (legal entity and tax registration ready: monotributo or equivalent, verify with an accountant). Log partner results in `enzymelens/validation/raw/` with checksums; no analysis.
- **Sun:** L-Sun only if the client agrees; otherwise a post on "what I learned in five discovery calls" with no names.
- **Gate 28 passes when:** first invoice sent (any amount). If not, week 29 is a second selling week.
- **Only one thing:** the proposal sent.

**Learn it.**
- ▶ **Core** (no study block; 30 minutes max)
  - Your own v4 Part 4 pricing hypotheses and Part 3.3 culture filter.
  - AFIP/ARCA guidance on issuing invoices as an independent professional in Argentina (verify current rules with an accountant; this plan does not give tax advice).
- ＋ **Deeper**
  - *Getting to Yes* (Fisher, Ury & Patton; book, paid) — principled negotiation, if price is contested.
  - Stanford eCorner talks on first customers (search).

**Run it.**

| Tool | ★ | Run this |
|---|---|---|
| Your proposal and invoice templates | — | Keep them in a private `business/` repo; version every proposal |
| `enzymelens/validation/` | — | `sha256sum` every raw result file; commit the checksums, not confidential data if the partner prefers |

**NotebookLM kit · `LABLOOP-W28 · First invoice gate`** (gate week: add Video Overview)
- **Load (8):** (1) this chapter; (2) scoping sheets from the calls (anonymised); (3) proposal; (4) v4 Part 4; (5) *The Mom Test* notes; (6) culture filter; (7) week 27 notebook Briefing doc; (8) `OPPORTUNITIES.csv` as text.
- **Focus line:** `Focus: turning discovery calls into a scoped, fixed-price proposal and closing the first paid engagement calmly.`
- **Custom question:** `From my call notes, which prospect has the clearest pain, budget and decision-maker? Draft the one-page proposal for that one.`

**Mentor check.** *"Someone paid me for measuring their loop."*

---

### Week 29 · Audit delivery 1

**Act VI · Threads: $, R · Next gate: 32**

**Welcome.** A good doctor's first visit is mostly listening and measuring: history, vitals, a few tests, and only then a diagnosis. The first audit week is that visit. You map the client's experimental loop from idea to decision, inventory every data source (paper notebooks, instrument exports, spreadsheets, email), measure cycle time and data readiness, and run one instrument on their sample data to produce a first prototype. Your corporate delivery muscle is exactly right here: interviews, a process map, a RAID log, a readout. The scientific muscle adds the part consultants usually skip: numbers with definitions, and a note on how each was measured. If the sale slipped, this week is the second selling week; the protocol for that is v5 Part 11.

**Where this week sits.**
- ← **From:** week 28 sale; instruments from weeks 1-26.
- → **Feeds:** week 30 (readout and case study), week 32 (second audit or pilot proposal), week 37 (pilot delivery).
- ⟳ **Returns in:** the anonymised case study, your most persuasive commercial asset.

**Goals.**
- 🎯 **This week:** loop map, data inventory, cycle-time and readiness numbers, one prototype on the client's sample; client readout.
- 🏁 **Next milestone:** Gate 32.
- 🌟 **Final:** a repeatable audit method.

**The plan (from v5 Part 7, expanded; gate and output unchanged).**
> *v5 row:* 29 | Audit delivery 1 | Client's domain reading | Loop map, data inventory, cycle-time and readiness numbers, one prototype on their sample | Client readout
- **Goal:** the client sees their loop with numbers for the first time.
- **Tue (study):** the client's two most relevant papers or patents (P11) and their product category basics.
- **Build:** [ ] Mon: 3-4 interviews (founder, lab lead, whoever runs the instruments); loop map (swimlanes: design → run → measure → analyse → decide). [ ] Wed: data inventory (source, format, owner, frequency, readiness score 0-3) and cycle-time measurement from their own records. [ ] Fri: prototype (batchlens, trialkit or a FermLit-style extraction) on their sample; readout deck.
- **Sun:** no public post about the client. L-Sun on a general lesson if useful.
- **Only one thing:** the loop map with cycle-time numbers.
- **If the sale slipped:** run a second selling week (week 27 plan, new 10 names) and push delivery by one week.

**Learn it.**
- ▶ **Core**
  - The client's papers/patents via P11 (60 min).
  - Coursera *Process Mining: Data Science in Action* (Eindhoven), week 1 lecture on event logs (30 min) — cycle time from timestamps is process mining in miniature.
- ＋ **Deeper**
  - Wil van der Aalst, *Process Mining: Data Science in Action* (book, paid).
  - NIST e-Handbook ch. 6.1 (process monitoring intro) as a vocabulary refresher.

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| pola-rs/polars | ★ 39.7k | Compute cycle times from the client's timestamps (experiment start → decision) with group-bys |
| docling-project/docling | ★ 66.3k | If their records are PDFs, convert a sample locally (on-prem; no data leaves your machine) |
| batchlens / trialkit (yours) | — | Run the relevant instrument on their sample in a private branch; never push client data |

**NotebookLM kit · `LABLOOP-W29 · Audit delivery 1`** (private notebook; client data only with written consent, otherwise anonymised summaries)
- **Load (8):** (1) this chapter; (2) client's public papers/patents; (3) anonymised interview notes; (4) loop map description; (5) data inventory table; (6) cycle-time numbers; (7) prototype output summary; (8) Process Mining lecture 1.
- **Focus line:** `Focus: mapping an experimental loop, scoring data readiness, measuring cycle time, and choosing the automation candidates with most days saved.`
- **Custom question:** `From the inventory and cycle times, rank three automation candidates by days saved per cycle divided by effort, and state each assumption explicitly.`

**Mentor check.** *"I can show a founder where their loop loses days, with numbers from their own records."*

---

### Week 30 · Audit delivery 2; Paper 3 results

**Act VI · Threads: M, $, EL · Next gate: 32**

**Welcome.** Two deliveries meet this week, and they rhyme. For the client, you close the audit with a readout and an anonymised case study (with permission). For the partner, you open the envelope: the wet-lab results for your candidates, analysed exactly as pre-registered in week 24. Small numbers of samples need special care: with three replicates, one outlier can create or destroy a "finding". Think of a restaurant with three reviews: one angry customer changes the rating completely, so you look at every review, not just the average. You report effect sizes with intervals, show every data point, and write down what worked, what did not, and what surprised you. A negative result is still a result; it is also what makes the positive ones credible.

**Where this week sits.**
- ← **From:** week 24 pre-registration, week 28 raw results, week 29 audit.
- → **Feeds:** week 31 (Paper 3 draft), week 32 (preprint + case study), week 38 (talk), week 39 (fund package uses the case study).
- ⟳ **Returns in:** L4 co-authored and the commercial proof of Act VI.

**Goals.**
- 🎯 **This week:** audit final readout; anonymised case study drafted; partner validation analysed per plan; Paper 3 figures.
- 🏁 **Next milestone:** Gate 32.
- 🌟 **Final:** Paper 3 with real validation and a public case study.

**The plan (from v5 Part 7, expanded; gate and output unchanged).**
> *v5 row:* 30 | Audit delivery 2; Paper 3 results | Statistics for validation results (small n) | Analysis of the partner's validation; figures | Anonymized case study drafted
- **Goal:** two honest reports: one for the client, one for science.
- **Tue (study):** statistics for small n: estimation plots, bootstrap limits with n = 3, why to show every point.
- **Build:** [ ] Mon: audit readout delivered; three candidates with days saved; client feedback captured; case-study permission asked. [ ] Wed: validation analysis exactly as in `PREREGISTRATION.md`; deviations from the plan listed explicitly. Done when every pre-registered analysis has a result. [ ] Fri: Paper 3 figures (predicted vs measured, per candidate, all replicates shown); case study draft (anonymised).
- **Sun:** L-Sun only with permissions; otherwise "what a pre-registration felt like when the data arrived".
- **Only one thing:** the pre-registered analysis run.

**Learn it.**
- ▶ **Core**
  - Tue · Ho et al. (2019), "Moving beyond P values: data analysis with estimation graphics", *Nature Methods* 16:565-566 (15 min); Cumming (2014) "The new statistics" (carry over), sections on small samples (30 min); StatQuest "Statistical Power" (carry over).
  - Weissgerber et al. (2015), "Beyond bar and line graphs: time for a new data presentation paradigm", *PLoS Biology* 13(4):e1002128 (20 min).
- ＋ **Deeper**
  - Vaux, Fidler & Cumming (2012), "Replicates and repeats—what is the difference and is it significant?", *EMBO Reports* 13:291-296.
  - Coursera *Introduction to Statistics* (Stanford), module on small-sample inference.

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| ACCLAB/DABEST-python | ★ 413 (the library from Ho et al. 2019) | Estimation plots (Gardner-Altman) for measured activity vs control; the library from Ho et al. 2019 |
| mwaskom/seaborn | ★ 14.0k | Strip plots with every replicate over the predicted ranking |
| arviz-devs/arviz | ★ 1.9k | If you fit a small Bayesian model, plot posteriors per candidate |
| lmfit/lmfit-py | ★ 1.2k | Run the frozen week 24 analysis script on the real data, unchanged |

**NotebookLM kit · `LABLOOP-W30 · Audit readout and validation results`**
- **Load (10):** (1) this chapter; (2) `PREREGISTRATION.md`; (3) Ho et al. 2019; (4) Weissgerber 2015; (5) Vaux 2012; (6) validation results summary (as permitted); (7) Paper 3 figure captions; (8) audit readout (anonymised); (9) case study draft; (10) week 24 Briefing doc.
- **Focus line:** `Focus: analysing small-n wet-lab validation honestly with estimation plots, reporting deviations from pre-registration, and writing an anonymised consulting case study.`
- **Custom question:** `Compare my analysis with the pre-registration line by line. Where did I deviate, and does any deviation change the conclusion?`

**Mentor check.** *"I can show whether my predictions held in a real lab, with every replicate visible, and I delivered a client readout the same week."*

---

### Week 31 · Write Paper 3

**Act VI · Threads: EL, R · Next gate: 32**

**Welcome.** Writing with a co-author is like cooking in someone else's kitchen: you bring your recipe, they know where everything is, and the dish is better when you respect their habits. Your partner knows what biologists expect in a validation section, which claims will embarrass them in front of peers, and which journals their community reads. You know how to make every number reproducible. This week you use the same writing process as Papers 1 and 2 (skeleton, P10, human read) and add the co-author loop: draft to them, comments back, their lab reads it. Expect their edits to feel conservative. They are protecting a reputation built over decades; that caution is now protecting yours too.

**Where this week sits.**
- ← **From:** weeks 21-30 (model, candidates, validation); weeks 14-18 writing craft.
- → **Feeds:** week 32 (preprint), week 36 (journal choice), week 38 (talk), Part 9 headline ("co-author with [lab]").
- ⟳ **Returns in:** the relationship: co-authors refer clients, students and future papers.

**Goals.**
- 🎯 **This week:** Paper 3 draft with co-author comments incorporated; P10 run; co-author's lab has read it.
- 🏁 **Next milestone:** Gate 32.
- 🌟 **Final:** L4 co-authored.

**The plan (from v5 Part 7, expanded; gate and output unchanged).**
> *v5 row:* 31 | Write Paper 3 | Research craft refresher | Draft with co-author; Prompt P10; co-author's lab reads it | Draft to co-author
- **Goal:** a draft both authors are proud of.
- **Tue (study):** research craft refresher (Gopen & Swan re-read; Writing in the Sciences module on discussion sections).
- **Build:** [ ] Mon: skeleton (P9) with the partner's section preferences; Methods (computational) by you, Methods (wet lab) by the partner. [ ] Wed: Results and Discussion draft; P10 twice. [ ] Fri: draft to co-author; CRediT statement; data and code availability agreed.
- **Sun:** no public post about unpublished results.
- **Only one thing:** draft sent to the co-author.

**Learn it.**
- ▶ **Core**
  - Tue · Gopen & Swan (1990) re-read (30 min); Coursera *Writing in the Sciences* (Stanford), module on "Discussion" and "Abstract" (45 min).
- ＋ **Deeper**
  - Target journal author guidelines chosen with the partner (e.g., a protein engineering or applied microbiology journal; verify fit and open-access policy).
  - Stanford CS197 lecture on co-authoring and feedback (verify).
  - Whitesides 2004 (carry over).

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| quarto-dev/quarto-cli | ★ 6.0k | Paper 3 manuscript; export to .docx for a co-author who uses Word track changes |
| enzymelens (yours) | — | `make paper` from a clean clone; tag `paper-3-draft-1` |
| Future-House/paper-qa | ★ 9.2k | Check the Discussion's claims about prior work against the PDFs you cite (verify manually) |

**NotebookLM kit · `LABLOOP-W31 · Paper 3 co-author draft`**
- **Load (10):** (1) this chapter; (2) Paper 3 draft; (3) co-author comments; (4) Gopen & Swan; (5) target journal guidelines; (6) TemStaPro and FLIP papers (related work); (7) Yang/Wu/Arnold 2019; (8) validation analysis summary; (9) P10 output; (10) CRediT statement.
- **Focus line:** `Focus: writing an applied protein-ML paper with wet-lab validation alongside a biologist co-author, and making claims that match the evidence.`
- **Custom question:** `Find every sentence in my Discussion that claims more than the validation shows, and propose a calibrated rewrite.`

**Mentor check.** *"A biologist and I wrote the same paper, and both of us would defend every sentence."*

---

### Week 32 · GATE 32: Paper 3 preprint (L4 co-authored)

**Act VI (finale) · Threads: EL, $, R · GATE 32**

**Welcome.** The local guide signs the map. Paper 3 goes to bioRxiv with your co-author's name next to yours, and the case study goes public with your client's permission. These two artifacts together say something no certificate can: *scientists trust my method, and companies pay for it.* You also send a second audit or pilot proposal, because the best time to sell the second engagement is the week the first one becomes a story. And you refresh your career assets: headline, site, applications with three preprints. Look at your gauge in `PROGRESS.md`. You are at L4. Twelve months ago you were at L0.

**Where this week sits.**
- ← **From:** Acts V and VI.
- → **Feeds:** week 36 (journal decisions), week 37 (pilot or job), week 38 (talk), week 39 (fund package), week 40 (decision).
- ⟳ **Returns in:** every introduction from now on: "co-author with [lab]".

**Goals.**
- 🎯 **This week:** Paper 3 on bioRxiv with co-author; case study published with permission; second audit or pilot proposal sent; career assets refreshed.
- 🏁 **This gate (32):** L4 co-authored.
- 🌟 **Final:** L5 reviewed and invited (weeks 36-40).

**The plan (from v5 Part 7, expanded; gate and output unchanged).**
> *v5 row:* 32 | GATE 32: Paper 3 preprint (L4 co-authored) | none | bioRxiv submission with co-author; second audit or pilot proposal sent | **X3**. Case study published with client permission
- **Goal:** X3 exists and the practice has a second opportunity in motion.
- **Build:** [ ] Mon: final co-author sign-off; bioRxiv submission; code/data DOIs. [ ] Wed: case study published (LinkedIn article + site); second audit or pilot proposal sent to the week 29 client or the best week 27 prospect. [ ] Fri: headline and site updated; 5 new applications or fractional-role conversations with three preprints (Part 9.4).
- **Sun:** L-Sun: the preprint, credited to both authors. **X3**. Gauge: **L4 (co-authored)**.
- **Gate 32 passes when:** bioRxiv DOI with co-author exists.
- **Only one thing:** the submission.

**Learn it.**
- ▶ **Core** (no study block)
  - bioRxiv submission guide (carry over); co-author's institution's preprint policy (ask).
- ＋ **Deeper**
  - Andrew Ng, *How to Build Your Career in AI* (free eBook from DeepLearning.AI; verify availability) — chapters on job search and portfolio.
  - Life Sci VC (Bruce Booth) posts on how biotech boards evaluate consultants and fractional executives (search).

**Run it.**

| Tool | ★ | Run this |
|---|---|---|
| Zenodo + GitHub (yours) | — | Release `enzymelens v1.0` with DOI; link from the preprint |
| marimo-team/marimo | ★ 22.7k | Interactive "candidate explorer" for the paper's supplementary material (public data only) |

**NotebookLM kit · `LABLOOP-W32 · Paper 3 gate and career refresh`** (gate week: add Video Overview + Slide Deck; then create `LABLOOP-ACT-VI`)
- **Load (12):** (1) this chapter; (2) Paper 3 preprint; (3) case study; (4) second proposal; (5) updated site text; (6) Papers 1-2 PDFs; (7) trialkit + batchlens READMEs; (8) 5 target role descriptions; (9) `PROGRESS.md` gauge; (10-12) `LABLOOP-ACT-IV/V` Briefing docs and week 30 Briefing doc.
- **Focus line:** `Focus: the full arc from first instrument to co-authored validated study and paid audit, told for founders, funds and hiring managers.`
- **Slide Deck prompt:** `Credentials deck, 8 slides: who I am in one line, three preprints, three instruments, one case study, the Loop Audit, how I work (fixed price, on-prem option), references, contact.`
- **Custom question:** `Write three versions of my 2-minute introduction: for a biotech founder, for a fund partner, for a hiring manager at a scientific-data company.`

**Mentor check.** *"I am a co-author with a scientist and a paid consultant with a public case study."*

---


# ACT VII · The lab's memory (weeks 33-35)

> **Act VII story.** Every lab has a person everyone asks: "where is the protocol we used in March?", "who changed the buffer recipe?", "did we ever see this contamination before?". When that person goes on holiday, the lab loses its memory. Regulators have a stricter version of the same worry: in a GMP facility, if a record cannot show who did what, when, and why it changed, it legally did not happen. Act VII builds the archive and the librarian: documents extracted locally (nothing leaves the building), written into a real open-source electronic lab notebook with a tamper-evident audit trail, and a question-answering assistant that cites the exact paragraph and is measured against human-labelled answers. It is FermLit's method turned inward, onto a lab's own documents, and it is the on-prem argument that IP-sensitive biotech companies care about most.

---

### Week 33 · Instrument 3 labdocs, part 1

**Act VII · Threads: M, $, R · Next gate: 36**

**Welcome.** Think of a bank vault ledger from the days before computers: each line written in ink, numbered, dated, signed, and never erased; a mistake is struck through with a single line and initialled. That is ALCOA+ (Attributable, Legible, Contemporaneous, Original, Accurate, plus complete, consistent, enduring, available), and 21 CFR Part 11 is the rule that says an electronic record must behave like that ledger. This week you read those rules (almost nobody in the startups has), install eLabFTW on the Debian box, and build the first pipe: scanned PDF → local vision model → structured experiment → eLabFTW entry → a hash-chained audit log, where each entry contains the fingerprint of the previous one, so any silent edit breaks the chain like a torn page in a numbered ledger.

**Where this week sits.**
- ← **From:** v4 Part 5.3 (labdocs design), week 5-7 (extraction with provenance), week 15 (GROBID/Docling), your local-model setup from week 0.
- → **Feeds:** week 34 (retrieval over the same records), week 35 (deviation drafting + benchmark), archetype D companies (traceability/QMS), week 37 (pilot option), week 39 (fund package).
- ⟳ **Returns in:** any client with regulatory pressure.

**Goals.**
- 🎯 **This week:** eLabFTW running; ingestion pipeline with a local VLM; hash-chained audit log; 200 synthetic SOPs/protocols; L-Sun.
- 🏁 **Next milestone:** Gate 36 (journal + JOSS submissions).
- 🌟 **Final:** `labdocs v0.1` and the on-prem benchmark table.

**The plan (from v5 Part 7, expanded; gate and output unchanged).**
> *v5 row:* 33 | Instrument 3 labdocs, part 1 | 21 CFR Part 11 + ALCOA+ (v4 catalog 15) | eLabFTW on the Debian box; local VLM extraction; audit log (v4 week 14) | L-Sun
- **Goal:** a scanned page becomes an auditable ELN record without leaving the building.
- **Tue (study):** 21 CFR Part 11 (sections 11.10 and 11.50), FDA "Part 11, Electronic Records; Electronic Signatures — Scope and Application" guidance (2003), MHRA "GxP Data Integrity Guidance and Definitions" (2018) on ALCOA+.
- **Build:** [ ] Mon: eLabFTW via Docker on the Debian box; API key; `elabapi-python` creates an experiment. Done when a test experiment appears in the UI. [ ] Wed: 200 synthetic SOPs/protocols from public templates (never client data); local VLM (Qwen2.5-VL 3B or Gemma 3 4B on the 3050) extracts fields to JSON with page/region provenance. Done when 20 pages extract and are scored against hand labels. [ ] Fri: write to eLabFTW + append-only audit log with SHA-256 chaining; verification command detects a tampered entry. Done when `labdocs verify` fails on a modified record.
- **Sun:** L-Sun "What ALCOA+ means in code".
- **Only one thing:** the tamper-detecting audit log.

**Learn it.**
- ▶ **Core**
  - Tue · 21 CFR Part 11, §11.10 and §11.50 (ecfr.gov) (20 min); FDA Part 11 Scope and Application guidance (2003), sections III-IV (25 min); MHRA GxP Data Integrity Guidance (2018), definitions section (30 min).
- ＋ **Deeper**
  - FDA, "Data Integrity and Compliance With Drug CGMP: Questions and Answers" (guidance, December 2018).
  - FDA "Computer Software Assurance for Production and Quality System Software" guidance (verify final status and date).
  - ICH E6(R3) Good Clinical Practice, data governance sections (only if a clinical client appears).
  - eLabFTW documentation — doc.elabftw.net (installation and API).
  - Coursera *Research Data Management and Sharing* (UNC/Edinburgh), modules on data storage and preservation.

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| elabftw/elabftw | ★ 1.4k (the most used open-source ELN) | Docker install per docs; create a team, a template, an experiment |
| elabftw/elabapi-python | ★ 44 (niche; official client) | Create and update an experiment from Python; read the revision history eLabFTW keeps |
| docling-project/docling | ★ 66.3k | Convert 10 synthetic SOP PDFs; compare layout/table fidelity with the VLM |
| datalab-to/marker | ★ 39.7k | Same 10 PDFs; note speed on the 3050 |
| ollama/ollama | ★ 180.8k | Run a vision model (e.g., `qwen2.5vl:3b` or `gemma3:4b`) on one scanned page; measure seconds per page |

**NotebookLM kit · `LABLOOP-W33 · labdocs ingestion and traceability`**
- **Load (10):** (1) this chapter; (2) v4 Part 5.3; (3) 21 CFR Part 11 text (ecfr page); (4) FDA Part 11 scope guidance PDF; (5) MHRA data integrity guidance PDF; (6) FDA data integrity Q&A 2018 PDF; (7) eLabFTW docs (API page); (8) Docling GitHub page; (9) your audit-log design note; (10) extraction scores on 20 pages.
- **Focus line:** `Focus: ALCOA+ and 21 CFR Part 11 explained as an ink ledger, how an ELN keeps an audit trail, and local document extraction for IP-sensitive labs.`
- **Studio extra (Infographic):** `ALCOA+ as nine icons with one-line meanings and, for each, the line of labdocs code or feature that implements it.`
- **Custom question:** `Which ALCOA+ principles does my current pipeline not yet satisfy, and what is the smallest change for each?`

**Mentor check.** *"I can explain ALCOA+ to a QA manager and prove a tampered record is detected."*

---

### Week 34 · labdocs, part 2

**Act VII · Threads: M, $, R · Next gate: 36**

**Welcome.** A librarian who answers confidently but sends you to the wrong shelf is worse than no librarian. This week you build "Ask the lab": local embeddings over the lab's records, answers that cite the exact protocol paragraph, and, as always, a measurement. You write 60 questions with human-labelled answers (your gold set again, now for retrieval), and you calibrate an LLM judge against your own labels so you know how much to trust automatic scoring. Retrieval has two failure modes worth remembering with an analogy: the librarian brings the wrong book (retrieval failure) or brings the right book and misreads it (generation failure). You measure them separately, because they are fixed in different places.

**Where this week sits.**
- ← **From:** week 33 records; weeks 6-7 and 10 (gold sets, agreement); week 14 (calibrated judge).
- → **Feeds:** week 35 (deviation drafting uses retrieval), week 37 (pilot option for archetype D), week 38 (talk example).
- ⟳ **Returns in:** every "chat with your documents" request a client makes; you will be the one who asks "measured against what?".

**Goals.**
- 🎯 **This week:** "Ask the lab" with citations; 60 labelled Q&A; retrieval and generation metrics separately; judge-human agreement number.
- 🏁 **Next milestone:** Gate 36.
- 🌟 **Final:** the calibrated-evaluation signature applied to regulated documents.

**The plan (from v5 Part 7, expanded; gate and output unchanged).**
> *v5 row:* 34 | labdocs, part 2 | Retrieval evaluation | "Ask the lab" with citations; 60 labeled Q&A; judge agreement (v4 week 15) | L-Sun
- **Goal:** a lab assistant whose trustworthiness is a number.
- **Tue (study):** retrieval-augmented generation and its evaluation (Lewis et al. 2020; RAGAS paper; your Vara notes).
- **Build:** [ ] Mon: chunking with section/paragraph IDs; local embeddings (bge-m3) on the Debian box; retrieval with citations. Done when every answer shows source paragraph IDs. [ ] Wed: 60 questions with human answers and gold source paragraphs; recall@k for retrieval; answer correctness separately. [ ] Fri: LLM judge vs human labels on the 60; agreement and the cases where the judge is wrong. Done when the agreement number is in the README.
- **Sun:** L-Sun "Judge-human agreement: X%. Everything else depends on that".
- **Only one thing:** the 60 labelled Q&A.

**Learn it.**
- ▶ **Core**
  - Tue · Lewis et al. (2020), "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks", NeurIPS, abstract + figure 1 (15 min); Es et al. (2023), "RAGAS: Automated Evaluation of Retrieval Augmented Generation", arXiv:2309.15217 (30 min); Stanford CS224N lecture on retrieval/RAG (web.stanford.edu/class/cs224n; verify current lecture list) (30 min).
- ＋ **Deeper**
  - Anthropic, "Introducing Contextual Retrieval" (engineering blog, 2024).
  - Thakur et al. (2021), "BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models", NeurIPS Datasets and Benchmarks.
  - Faysse et al. (2024), "ColPali: Efficient Document Retrieval with Vision Language Models", arXiv:2407.01449 — retrieval directly on page images.
  - Zheng et al. 2023 LLM-as-a-judge (carry over).

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| vibrantlabsai/ragas | ★ 15.7k | Compute context recall and faithfulness on 10 of your Q&A; compare with your human labels before trusting it |
| run-llama/llama_index | ★ 52.1k | Only as a reference implementation of chunking and citation; keep your own pipeline small |
| illuin-tech/colpali | ★ 2.8k | Try page-image retrieval on 20 scanned SOP pages; compare recall@5 with text chunks |
| UKGovernmentBEIS/inspect_ai | ★ 2.8k | Wrap the 60 Q&A as an eval with a model-graded scorer and your human labels |
| ggml-org/llama.cpp | ★ 128.1k | Serve the embedding model locally on the mini PC for the on-prem story |

**NotebookLM kit · `LABLOOP-W34 · labdocs retrieval evaluation`**
- **Load (10):** (1) this chapter; (2) Lewis 2020; (3) RAGAS paper; (4) BEIR paper abstract; (5) ColPali paper; (6) Anthropic contextual retrieval post; (7) Zheng 2023; (8) your 60 Q&A (synthetic docs only); (9) recall@k and correctness tables; (10) judge-human disagreement cases.
- **Focus line:** `Focus: retrieval-augmented generation with citations, separating retrieval failures from generation failures, and calibrating an LLM judge against human labels.`
- **Custom question:** `Classify each judge-human disagreement as judge too lenient, judge too strict, or ambiguous question. What rule would fix the largest group?`

**Mentor check.** *"I can tell a lab exactly how often their assistant cites the right paragraph and how much to trust the automatic scores."*

---

### Week 35 · labdocs, part 3

**Act VII (finale) · Threads: M, BL, $, R · Next gate: 36**

**Welcome.** This week the instruments start talking to each other. When batchlens raises an alarm (week 3), a QA person must write a deviation report: what happened, which SOP applies, what the corrective action is. labdocs drafts it: alarm + relevant SOP paragraph → draft → human approval → signed record in eLabFTW. The human stays in the loop, like a pilot who reviews the autopilot's plan before accepting it. Then the argument that sells labdocs to IP-sensitive companies: a benchmark table of what runs on a USD 300 mini PC, what needs your 3050, and what needs the cloud, with accuracy and cost per 1,000 pages. You also skim privacy basics, because the first question from any serious client will be "where does our data go?".

**Where this week sits.**
- ← **From:** weeks 3-4 (alarms), weeks 33-34 (records, retrieval).
- → **Feeds:** week 36 (JOSS options), week 37 (pilot), week 38 (talk demo), week 39 (fund package).
- ⟳ **Returns in:** every on-prem conversation.

**Goals.**
- 🎯 **This week:** deviation drafting with human approval; on-prem benchmark table; `labdocs v0.1` + Z6; M5 deep dive.
- 🏁 **Next milestone:** Gate 36.
- 🌟 **Final:** three instruments with DOIs, one of them regulated-ready in spirit.

**The plan (from v5 Part 7, expanded; gate and output unchanged).**
> *v5 row:* 35 | labdocs, part 3 | Privacy basics | Deviation drafting + on-prem benchmark table (v4 weeks 16-17) | `labdocs v0.1`, Z6, M5 deep dive
- **Goal:** alarm-to-signed-deviation in minutes, and a hardware/cost table.
- **Tue (study):** privacy and data governance basics: GDPR principles (Art. 5), Argentina Ley 25.326 overview, HIPAA summary (only concepts).
- **Build:** [ ] Mon: deviation drafting: batchlens alarm JSON + retrieved SOP → draft (CAPA fields) → approval step → eLabFTW record with audit entry. Done when one end-to-end deviation is signed. [ ] Wed: benchmark on 200 synthetic pages: mini PC vs 3050 vs one cloud model: accuracy per field, seconds/page, cost per 1,000 pages. Done when `make bench` regenerates the table. [ ] Fri: Docker compose of the stack; release `labdocs v0.1` + Zenodo DOI; M5 draft.
- **Sun:** **M5** deep dive "A lab's AI on a USD 300 computer: what runs where, and how well". Z6.
- **Only one thing:** the benchmark table.

**Learn it.**
- ▶ **Core**
  - Tue · GDPR Article 5 (principles) and Article 25 (privacy by design) (20 min); Argentina's Ley 25.326 summary on argentina.gob.ar (verify page) (20 min); HHS HIPAA "Summary of the Privacy Rule" (hhs.gov) (20 min).
- ＋ **Deeper**
  - NIST Privacy Framework overview (nist.gov).
  - Stanford HAI *AI Index* report, chapter on responsible AI (for context in M5).
  - v4 Part 5.4 (phase-2 backlog) — re-read to decide nothing yet.

**Run it.**

| Repo / notebook | ★ | Run this (20-40 min) |
|---|---|---|
| ggml-org/llama.cpp | ★ 128.1k | `llama-bench` on the mini PC and the 3050 for the same quantised model |
| ollama/ollama | ★ 180.8k | Same model via Ollama; record tokens/s and memory |
| docling-project/docling / opendatalab/MinerU | ★ 66.3k / ★ 79.8k | Add one row per converter to the benchmark (accuracy on tables, seconds/page) |
| allenai/olmocr | ★ 19.5k | Optional OCR row if a GPU run fits; mark hardware requirements honestly |

**NotebookLM kit · `LABLOOP-W35 · labdocs release and local benchmark`**
- **Load (9):** (1) this chapter; (2) GDPR Art. 5/25 text; (3) Ley 25.326 summary; (4) HIPAA summary; (5) v4 weeks 16-17 text; (6) benchmark table; (7) deviation workflow description; (8) M5 draft; (9) weeks 33-34 Briefing docs.
- **Focus line:** `Focus: human-in-the-loop deviation drafting, privacy-by-design basics, and the trade-offs of running document AI on a mini PC, a consumer GPU or the cloud.`
- **Studio extra (Infographic):** `Three-column comparison: mini PC, RTX 3050, cloud: what runs, accuracy, seconds per page, cost per 1,000 pages, where the data goes.`
- **Custom question:** `Write the privacy paragraph of my labdocs one-pager for an Argentine biotech handling proprietary strains, without overpromising compliance.`

**Mentor check.** *"I can show an alarm becoming a signed deviation, and a table of what a lab can run without sending data anywhere."*

---

# ACT VIII · Peer review and the crossroads (weeks 36-40)

> **Act VIII story.** The expedition comes home. The maps go to the geographical society, where reviewers you have never met will check them. You give the lecture. A fund asks whether you can survey all its territories at once. And at the end, standing at the crossroads, you decide where to settle: your own office with long-term clients, a salaried seat at a map-making company, or both. The decision is written down, because a decision in your head is a mood.

---

### Week 36 · GATE 36: journal + JOSS

**Act VIII · Threads: FL, BL, R · GATE 36**

**Welcome.** Preprints are like a published speech; journals are like the speech after the fact-checkers read it. Some doors (grant panels, pharma QA departments, certain hiring committees) still open only for peer-reviewed work. This week you send Paper 1 to a data journal, Paper 2 to a process-engineering or bioengineering journal, and batchlens to JOSS, where reviewers literally open GitHub issues on your repo. Choosing a journal is like choosing a conference to present at: the right audience matters more than the brand. Read each journal's scope, look at five recent papers to check you "sound like" them, and adapt the format. Then submit and let go; reviews take months, and the plan keeps moving.

**Where this week sits.**
- ← **From:** Papers 1-2 (weeks 16, 20), batchlens releases (weeks 4, 8, 17).
- → **Feeds:** week 40 (L5 path), your academic credibility for years; JOSS acceptance can arrive within weeks.
- ⟳ **Returns in:** revisions over the next months (plan 2 hours per revision round in Saturday blocks).

**Goals.**
- 🎯 **This week:** two journal submissions (one each for Paper 1 and 2) and one JOSS submission, all logged.
- 🏁 **This gate (36):** submissions logged.
- 🌟 **Final:** L5 reviewed and invited.

**The plan (from v5 Part 7, expanded; gate and output unchanged).**
> *v5 row:* 36 | GATE 36: journal + JOSS | Journal author guides | Paper 1 to Scientific Data or a bioprocess journal; Paper 2 to Computers & Chemical Engineering or Biotechnology and Bioengineering (pick one each); JOSS submission for batchlens | Submissions logged
- **Goal:** three submissions in the system.
- **Tue (study):** author guides for the chosen journals; JOSS review criteria.
- **Build:** [ ] Mon: Paper 1 reformatted as a *Scientific Data* Data Descriptor (or a bioprocess journal's format); submitted. [ ] Wed: Paper 2 reformatted for *Computers & Chemical Engineering* or *Biotechnology and Bioengineering*; cover letter; submitted. [ ] Fri: JOSS `paper.md` (≤ 1,000 words), statement of need, tests, docs, community guidelines; submitted. Log all three in `PUBLICATIONS.csv`.
- **Sun:** L-Sun "Three submissions: what changes between a preprint and a journal paper".
- **Gate 36 passes when:** two journal submissions and one JOSS submission are logged.
- **Only one thing:** the JOSS submission (fastest review, most useful for instruments).

**Learn it.**
- ▶ **Core**
  - Tue · *Scientific Data* submission guidelines for Data Descriptors (nature.com/sdata) (20 min); JOSS "Review criteria" and "Submitting a paper" (joss.theoj.org) (25 min); author guides of the chosen Paper 2 journal (20 min).
- ＋ **Deeper**
  - Coursera *Writing in the Sciences* module on "The publication process" and cover letters.
  - Smith et al. (2018), "Journal of Open Source Software (JOSS): design and first-year review", *PeerJ Computer Science* 4:e147.
  - COPE (Committee on Publication Ethics) guidance on preprints and duplicate publication (publicationethics.org).

**Run it.**

| Repo | ★ | Run this (20-40 min) |
|---|---|---|
| openjournals/joss | ★ 1.9k | Read the docs folder on paper format; open two recently accepted JOSS review issues for bioprocess or scientific tools to see what reviewers ask |
| batchlens (yours) | — | Check the JOSS checklist: license, tests run in CI, docs, contribution guide, example usage |
| quarto-dev/quarto-cli | ★ 6.0k | Export to the journal's LaTeX template where available |

**NotebookLM kit · `LABLOOP-W36 · Journal and JOSS submissions`** (gate week: add Video Overview)
- **Load (10):** (1) this chapter; (2) *Scientific Data* guidelines page; (3) Paper 2 journal guide page; (4) JOSS review criteria page; (5) Smith 2018 JOSS paper; (6) Paper 1 PDF; (7) Paper 2 PDF; (8) batchlens README; (9) JOSS `paper.md` draft; (10) cover letters.
- **Focus line:** `Focus: moving from preprint to peer-reviewed journal, choosing venues by audience, and what JOSS reviewers check in research software.`
- **Custom question:** `Compare my Paper 1 with the Scientific Data Data Descriptor template. What must be moved, cut or added?`

**Mentor check.** *"My work is in peer review at three venues, and I know what each reviewer will look for."*

---

### Week 37 · Pilot delivery or job process

**Act VIII · Threads: $, R · Next gate: 40**

**Welcome.** Two roads, both good, and you may walk them in parallel. On the consulting road, the audit client (or a new one) runs a 6-8 week pilot: one instrument adapted to their data, running in their lab, with a measured before and after. On the employee road, scientific-data companies interview you with three preprints on the table. Your 25 years matter here more than anywhere else in the plan: you know how to run a weekly client readout, and you know how to read an interviewer. Treat interviews like a two-way audit: the culture filter from v4 is your checklist, and your Sumble friend's calm culture is the reference point. You are no longer asking "will they take me?"; you are asking "is this the right lab for my next ten years?".

**Where this week sits.**
- ← **From:** week 32 proposal, week 19/32 applications, v4 Part 3.3 culture filter.
- → **Feeds:** week 39 (fund package can bundle pilots), week 40 (decision with real options).
- ⟳ **Returns in:** income and offers in the final retro.

**Goals.**
- 🎯 **This week:** a pilot running with a weekly readout **or** an interview process active with at least one scientific-data company.
- 🏁 **Next milestone:** Gate 40.
- 🌟 **Final:** a real choice at week 40, not a hypothetical one.

**The plan (from v5 Part 7, expanded; gate and output unchanged).**
> *v5 row:* 37 | Pilot delivery or job process | Client or interview prep | Pilot (6-8 weeks) running, or interviews at scientific-data companies with three preprints | Weekly client readout
- **Goal:** one concrete path is moving.
- **Tue (study):** client domain reading or interview preparation (system design for scientific data, your papers as case studies).
- **Build:** [ ] Mon: pilot kickoff (success metric agreed, baseline measured) or interview prep (3 STAR stories per preprint). [ ] Wed: pilot data pipeline on client infrastructure (on-prem if needed) or technical interview practice. [ ] Fri: first weekly readout (one page: metric, change, blockers, next week) or interview debrief with the culture filter scored.
- **Sun:** private reflection in `LOG.md`; public post only about general lessons.
- **Only one thing:** the first weekly readout (or the culture-filter score after the first interview).

**Learn it.**
- ▶ **Core**
  - Tue · Chip Huyen, *ML Interviews Book* (free) — chapters relevant to the role (45 min); v4 Part 3.3 culture filter (10 min).
  - For pilots: re-read your audit readout and the client's success metric definition (20 min).
- ＋ **Deeper**
  - Chip Huyen, *Designing Machine Learning Systems* (book, paid) — chapters on data and deployment for the system-design interview.
  - *The Trusted Advisor* (carry over) for running a pilot relationship.
  - Company engineering/science blogs of the targets (DataHow, TetraScience, Benchling, Novasign; verify).

**Run it.**

| Repo | ★ | Run this (20-40 min) |
|---|---|---|
| chiphuyen/ml-interviews-book | ★ 4.8k | Practise 5 questions aloud, record, compare with the book's guidance |
| chiphuyen/aie-book | ★ 17.4k | Skim the resources list for evaluation and deployment topics likely in a scientific-data interview |
| marimo-team/marimo | ★ 22.7k | A one-page interactive pilot dashboard for the weekly readout (client data stays on their machine) |

**NotebookLM kit · `LABLOOP-W37 · Pilot or job process`** (private)
- **Load (9):** (1) this chapter; (2) v4 Part 3.3 culture filter; (3) pilot success metric and baseline (anonymised) **or** job descriptions; (4) three preprints; (5) instrument READMEs; (6) ML Interviews Book chapter pages; (7) interview notes; (8) weekly readout; (9) week 32 credentials deck.
- **Focus line:** `Focus: running a measured pilot with weekly readouts, preparing for scientific-data company interviews with my own papers, and evaluating culture fit.`
- **Custom question:** `Act as a hiring manager at a bioprocess-AI company. Interview me for 10 exchanges about Paper 2, then score me on depth, honesty about limitations and communication.`

**Mentor check.** *"I have a real option in motion, and I am evaluating it as carefully as it is evaluating me."*

---

### Week 38 · Talk delivered

**Act VIII · Threads: R · Next gate: 40**

**Welcome.** Scientists often learn more about a researcher from a 20-minute talk than from their papers, because a talk shows how you think when there is no edit button. Patrick Winston's rule is the one to carry: start with an empowerment promise ("by the end you will know..."), cycle back to the key idea several times, and end with a contribution, not "thank you, questions?". Your talk, proposed in week 20, tells one story: two thousand fermentation papers and a hundred open batches, what the data says, and what that means for a lab in Buenos Aires or Rosario. Rehearse it three times with Prompt P3 and once with a human. Record it; the video becomes a permanent asset that works while you sleep.

**Where this week sits.**
- ← **From:** week 20 proposal, all three papers, instruments, case study.
- → **Feeds:** week 39 (fund package credibility), week 40 (L5 invited), future invitations.
- ⟳ **Returns in:** every conference invitation after this one.

**Goals.**
- 🎯 **This week:** 20-minute talk delivered; video posted.
- 🏁 **Next milestone:** Gate 40.
- 🌟 **Final:** L5 "invited".

**The plan (from v5 Part 7, expanded; gate and output unchanged).**
> *v5 row:* 38 | Talk delivered | Talk rehearsal (Prompt P3) | 20-minute talk at the event accepted in week 20 | Video posted
- **Goal:** the talk happens and exists as a video.
- **Tue (study):** rehearsal with P3 (audience persona: mixed biotech founders and scientists).
- **Build:** [ ] Mon: slide deck from the NotebookLM Slide Deck draft, rewritten by you: one idea per slide, figures from the papers. [ ] Wed: two timed rehearsals, recorded; cut to 18 minutes. [ ] Fri (or event day): deliver; collect three contacts and one question you could not answer (it goes to `PARKING.md` as Half 3 seed).
- **Sun:** video posted with a LinkedIn summary.
- **Only one thing:** the talk delivered.

**Learn it.**
- ▶ **Core**
  - Patrick Winston, "How to Speak" (MIT OCW / YouTube, ~1 h; re-watch the first 30 min).
  - Simon Peyton Jones, "How to give a great research talk" (Microsoft Research, YouTube).
- ＋ **Deeper**
  - Nancy Duarte, "The secret structure of great talks" (TEDx, YouTube).
  - Coursera *Introduction to Public Speaking* (University of Washington), weeks 2-3.

**Run it.**

| Tool | ★ | Run this |
|---|---|---|
| marimo-team/marimo | ★ 22.7k | One live demo cell (the learning-curve slider) as the talk's "wow" moment; have a screenshot fallback |
| quarto-dev/quarto-cli | ★ 6.0k | Optional reveal.js slides from Markdown so figures regenerate from code |

**NotebookLM kit · `LABLOOP-W38 · Scientific talk and proof`**
- **Load (9):** (1) this chapter; (2) Winston "How to Speak"; (3) Peyton Jones research talk; (4) three preprints; (5) case study; (6) week 20 talk proposal and deck; (7) rehearsal transcripts (P18); (8) event audience description; (9) `LABLOOP-ACT-VII` Briefing doc.
- **Focus line:** `Focus: a 20-minute research talk for biotech founders and scientists: one promise, one story, two findings, one method, limitations, and a contribution slide.`
- **Custom question:** `Listen to my rehearsal transcript. Where do I lose the audience, where do I use jargon without definition, and what is my actual contribution sentence?`

**Mentor check.** *"I stood in front of scientists and founders and explained my work clearly, and there is a video to prove it."*

---

### Week 39 · Fund package

**Act VIII · Threads: $, R · Next gate: 40**

**Welcome.** A single startup is a village; a fund's portfolio is a region. Funds like GridX care about every company in the portfolio hitting milestones with the runway they have. A portfolio data-readiness package, five mini audits plus one consolidated report, gives the fund something rare: a comparable, measured view of how fast each company's experimental loop turns, and what would speed it up. This is also where your program-management past becomes a direct asset: you have run multi-workstream programmes for large organisations for 25 years. Offer it only now, when you have case studies and preprints to show; a fund forwards weak material to its companies once, and never again.

**Where this week sits.**
- ← **From:** week 16 (preprint sent to GridX contact), week 27 (offer), weeks 29-30 and 32 (case study), v4 Part 4 (portfolio package pricing hypothesis).
- → **Feeds:** week 40 (pipeline in the decision), Half 3 (retainers).
- ⟳ **Returns in:** long-term deal flow; funds are repeat buyers.

**Goals.**
- 🎯 **This week:** portfolio data-readiness package proposed to one fund (GridX first).
- 🏁 **Next milestone:** Gate 40.
- 🌟 **Final:** a repeatable, higher-ticket offer.

**The plan (from v5 Part 7, expanded; gate and output unchanged).**
> *v5 row:* 39 | Fund package | none | Portfolio data-readiness package proposed to one fund (GridX first) | Proposal sent
- **Goal:** the proposal is sent.
- **Build:** [ ] Mon: package design: 5 mini audits (1 week each, standard template), consolidated report (loop cycle times, readiness scores, top automation candidates across the portfolio), pricing from v4 Part 4 updated with real audit effort. [ ] Wed: proposal (2 pages) with the case study, preprints and one anonymised sample page of the consolidated report. [ ] Fri: sent to the GridX contact; follow-up date logged.
- **Sun:** no public post about the fund; general post on "what a portfolio view of lab data reveals" only when real.
- **Only one thing:** the proposal sent.

**Learn it.**
- ▶ **Core** (no study block; 30 minutes)
  - v4 Part 4 (portfolio package) and Part 9.4 (funds).
  - Bruce Booth, *Life Sci VC* (lifescivc.com) — one post on portfolio support or milestones (search).
- ＋ **Deeper**
  - a16z Bio + Health podcast (search an episode on biotech operations or AI in R&D).
  - *Nature Biotechnology* "Bioentrepreneur" section articles on working with investors.

**Run it.**

| Tool | ★ | Run this |
|---|---|---|
| marimo-team/marimo | ★ 22.7k | Sample consolidated-report page from synthetic data: 5 companies, cycle time, readiness, days saved |
| pola-rs/polars / plotly/plotly.py | ★ 39.7k / ★ 18.8k | Generate the sample report figures reproducibly |

**NotebookLM kit · `LABLOOP-W39 · Fund portfolio offer`**
- **Load (8):** (1) this chapter; (2) v4 Part 4 + 9.4; (3) case study; (4) audit template; (5) sample consolidated page; (6) three preprint abstracts; (7) GridX public portfolio page (website); (8) Life Sci VC post.
- **Focus line:** `Focus: what a venture fund needs from a portfolio-wide data-readiness assessment and how to price and present it.`
- **Custom question:** `Act as a fund partner. What would make you forward this proposal to five portfolio companies, and what would make you ignore it?`

**Mentor check.** *"I can offer a fund a measured view of its whole portfolio's lab loops, backed by a real case study."*

---

### Week 40 · GATE 40: retro and decision

**Act VIII (finale) · Threads: all · GATE 40**

**Welcome.** Stanford's *Designing Your Life* course has a simple idea: you do not find your path by thinking harder; you find it by prototyping lives and then choosing among real options. That is what the last 40 weeks were, whether you noticed or not: prototypes of a scientist's life (papers), a consultant's life (audits), and an employee's life (interviews). This week you look at the evidence, not the mood. Gauge level. Income to date. Pipeline. Offers. The culture filter. Your energy log: which weeks made you feel alive and which drained you. Then you write the decision: consultant with retainers, employee, or both. You also write the public retrospective, because the people who followed your 40 weeks deserve to know how the story ends, and because your next clients and colleagues will read it. Then you rest. A cartographer who never stops walking never draws the map.

**Where this week sits.**
- ← **From:** everything.
- → **Feeds:** Half 3 (month 10+): retainers, fractional role, or a new job; L6 "used" as others adopt FermLit and batchlens.
- ⟳ **Returns in:** your next plan, which starts from evidence rather than from zero.

**Goals.**
- 🎯 **This week:** retro with gauge, income, pipeline and offers; written decision; public retro.
- 🏁 **This gate (40):** see the pass criteria below.
- 🌟 **Final:** you are, verifiably, an AI scientist for biotech with a practice or a role of your choosing.

**The plan (from v5 Part 7, expanded; gate and output unchanged).**
> *v5 row:* 40 | GATE 40: retro and decision | none | Gauge level, income to date, pipeline, offers. Decide: consultant with retainers, or employee, or both | Public retro
- **Goal:** a decision written down, backed by evidence.
- **Build:** [ ] Mon: evidence table: every gate, artifact link, DOI, date; gauge level; income; pipeline; offers; lost weeks. [ ] Wed: decision memo (one page): options, evidence for each, culture filter scores, energy log patterns, what you choose and what you stop. [ ] Fri: public retro (LinkedIn article + site): what worked, what failed, what you would tell someone starting week 0.
- **Sun:** public retro posted. Plan Half 3 in one page, no more.
- **Gate 40 passes when:** three preprints (one co-authored), two journal submissions, one JOSS submission, three instruments with DOIs, one talk delivered, at least one paid engagement delivered, and a decision written down.
- **Only one thing:** the decision memo.

**Learn it.**
- ▶ **Core**
  - Bill Burnett & Dave Evans, *Designing Your Life* (book, paid) — chapters on "odyssey plans" and "choosing well"; or Bill Burnett's TEDx talk "Designing Your Life" (YouTube, free) (30 min).
  - Richard Hamming, *You and Your Research* (final re-read; compare your week 0 notes) (30 min).
- ＋ **Deeper**
  - Uri Alon's talk and essay on "the cloud" in research (iBiology; verify).
  - Stanford *Designing Your Life* resources (designingyour.life) — odyssey plan worksheets.

**Run it.**

| Tool | ★ | Run this |
|---|---|---|
| Your `cockpit` repo | — | Script that counts blocks, lost weeks and gates from `LOG.md` and `PROGRESS.md`; the only dashboard you ever need |
| marimo-team/marimo | ★ 22.7k | One-page interactive "40 weeks" timeline for the public retro |

**NotebookLM kit · `LABLOOP-W40 · Final retrospective and career decision`** (gate week: add Video Overview; then create `LABLOOP-ACT-VIII` and `LABLOOP-40-WEEKS`)
- **Load (14):** (1) this chapter; (2) `PROGRESS.md`; (3) `LOG.md`; (4) evidence table; (5) decision memo draft; (6) culture filter scores; (7) week 0 and week 20 retros; (8) three preprints; (9) case study; (10) talk transcript; (11) Hamming transcript; (12) Designing Your Life TEDx; (13-14) `LABLOOP-ACT-VI/VII` Briefing docs.
- **Focus line:** `Focus: an evidence-based retrospective of a 40-week transition from program manager to AI scientist for biotech, and a clear career decision.`
- **Slide Deck prompt (optional):** `Ten slides telling the 40-week story in eight acts, one number per act, and the decision.`
- **Custom question:** `Using only the evidence in these sources, argue for each of the three options (consultant, employee, both) as strongly as possible. Then tell me which argument relies least on hope.`

**Mentor check.** *"I know what I built, what it proved, what people paid for, and where I am going next. It is written down."*

---

## PART 8 · The prompt pack (copy, paste, fill the brackets)

Use Claude for P1-P2, P5-P10, P12-P15; NotebookLM (sources loaded) for P3-P4 and P16-P18; either for P7, P11, P13.

**P1 · Kid explainer (study, every Tue/Thu when a concept blocks)**
```
Explain [CONCEPT] to me three times.
1. As if I were 10: one everyday object, no jargon, no equations.
2. As if I were a systems engineer with 25 years of experience: what problem it solves, what breaks without it, an analogy from infrastructure (queues, caches, budgets), never from brains.
3. As the textbook would: the equation, every symbol named in one line each, and a tiny worked example with numbers I can redo by hand.
Then: one question to check I understood, and wait for my answer.
```

**P2 · Concept-to-project bridge (before a build block)**
```
I am building [batchlens / fermlit / enzymelens]: [one-sentence description].
Here is what I just studied: [paste notes].
Give me the 3 concrete places where this applies today, ordered by impact over effort. For each: file and function to touch, 10 lines of pseudocode max, the metric that should move, and how I would know I applied it wrong. Skip anything that requires training a model from scratch or new tools.
```

**P3 · Socratic quiz (NotebookLM, Sunday or review weeks)**
```
Act as [a PI running a lab meeting / a staff ML scientist at a bioprocess-AI company]. Ask me 5 questions of increasing difficulty about [TOPIC], one at a time, waiting for my answer. After each answer, score my rigor 1-5 and say exactly what was missing. Call out vagueness without diplomacy. The last question must be a design case with a cost or data constraint.
```

**P4 · Gap detector (NotebookLM, Sunday, 20 min)**
```
Below is what I understood this week. Compare it against the sources in this notebook.
Return: (1) what I said correctly, one line each; (2) what I said wrong or imprecisely, with the exact quote from the source that corrects me; (3) what I did not mention and should have, ordered by importance; (4) the single thing to revisit on Monday.
MY SUMMARY: [paste]
```

**P5 · Practice scaffold (turns a lecture into one 90-minute build block)**
```
I have 90 minutes, a Python repo [NAME], and this lecture/chapter: [summary or link]. My project needs [the week's build task].
Write me a single-block plan: the one file to create or edit, the function signatures, the test that proves it works, and the "done when" line. No alternatives, no options, no "you could also". If the task does not fit in 90 minutes, cut scope and tell me what you cut.
```

**P6 · Staff-engineer code review (Fridays, before release)**
```
Review this code as a staff engineer at a scientific-software company who cares about reproducibility.
Report only: (1) anything that makes a result non-reproducible (seeds, data leakage across batches, hidden state); (2) anything a reviewer of a paper would flag; (3) the smallest change that fixes each. Ignore style.
[paste code or diff]
```

**P7 · Annotation guidelines writer (week 6)**
```
I am building a gold standard for extracting these fields from fermentation papers: [schema]. Write annotation guidelines: for each field, a one-line definition, the unit and normalization rule, two positive examples and two tricky negative examples (things that look like the field but are not), and what to do when the paper reports a range, several conditions, or nothing. End with a 10-item checklist an annotator runs before saving a paper.
```

**P8 · Related-work scout (with verification rule)**
```
I am writing a paper on [topic]. List the 15 most relevant prior works you know: title, first author, year, venue, one line on what they did, and one line on how mine differs. Mark each as [SURE] or [VERIFY]. I will check every one on Semantic Scholar before citing; do not invent titles. Then list 5 search queries I should run to find what you might have missed.
```

**P9 · Paper skeleton (weeks 8, 13, 18, 31)**
```
From my notes and results below, build the skeleton of a paper for [venue]. Structure: title with a measured number; abstract with the numbers; introduction (problem in one paragraph, gap, contributions as 3 bullets); related work headings; methods sections with the exact subsections; results with one heading per figure and the figure caption drafted; limitations; data and code availability. Mark [VERIFY] on every claim not supported by my numbers. Do not write prose I have not earned; leave bullet points where I still lack evidence.
NOTES AND RESULTS: [paste]
```

**P10 · Reviewer 2 (every draft)**
```
You are a skeptical reviewer for [venue] with expertise in [bioprocess modeling / information extraction / protein ML]. Read the draft and produce: (1) three reasons to reject, most damaging first; (2) every claim that lacks evidence or an interval; (3) every place where a method is not reproducible from the text; (4) the missing related work you suspect exists (mark [VERIFY]); (5) what would change your recommendation to accept. Be harsh and specific; quote the lines.
[paste draft]
```

**P11 · Founder paper digest (Thursday, 30 min)**
```
Digest this paper for me: (1) the question and why it matters, in two lines; (2) the method, in five lines a non-biologist can follow; (3) the main result with its numbers; (4) the data they generated and how they analyzed it (tools, statistics, spreadsheets or code?); (5) three questions I could ask the author that show I read it; (6) one honest sentence I could write to them that connects their data workflow to what I build, with no pitch.
[paste paper text or abstract + methods]
```

**P12 · Unblocker (when frozen at the start of a block)**
```
I am frozen. The task is: [TODAY.md line]. Ask me three questions, one at a time: what is the smallest visible thing I could produce in 10 minutes; what am I afraid will happen if I start; what would I tell a colleague to do. Then give me the first 3 commands or lines to type. Nothing else.
```

**P13 · Endorsement or co-authorship request (weeks 12-19, 21-24)**
```
Draft a short email to [NAME], author of [PAPER], from someone with 25 years in engineering and a first preprint in [topic]. Purpose: [arXiv endorsement for category X / feedback / co-authorship on a validation study]. Include: one specific thing from their paper I built on (I will fill it), the link to my preprint and DOI, exactly what I am asking and how long it takes them, and an easy no. Under 150 words. No flattery.
```

**P14 · Lost-week compressor (Sunday after a lost week)**
```
I lost this week. The plan for next week is: [paste week]. Cut it to the single build task and the single study unit that keep the next gate alive, and tell me what to skip permanently versus what to move. Give me the six TODAY.md lines.
```

**P15 · Week expander (for other models, Half 2)**
```
Expand week [N] of this master plan into the same format as Part 6: goal, Tue and Thu study with exact units and links, three build tasks with "done when", Sunday output, "only one thing". Do not change the gate, the output, or the rules in Part 0. Use only free resources and mark any link you are not sure of as (verify).
```

**P16 · Visual explainer brief (NotebookLM, when an infographic or mind map is not enough)**
```
Using only the sources in this notebook, design a one-page visual explanation of [CONCEPT] for someone with 25 years of engineering and program management experience and no biology degree.
Give me: (1) the central metaphor in one sentence; (2) 5-7 boxes in a left-to-right flow, each with a title of max 4 words, one number or equation, and the source it comes from; (3) the arrows between boxes labelled with the verb that connects them; (4) one "common misconception" callout with the source that corrects it; (5) a final arrow "where this goes next" pointing to [NEXT WEEK'S TOPIC].
Then generate the infographic with that structure.
```

**P17 · Connection map (NotebookLM, Sunday of gate weeks or when a week feels disconnected)**
```
I have pasted the Briefing docs of weeks [N-3] to [N] as sources. Build a connection map:
(1) for each week, the one concept and the one artifact it produced;
(2) every place where a later week reuses an earlier concept or artifact, as "week X → week Y: what is reused";
(3) the thread (Biology, AI method, batchlens, FermLit, enzymelens, money, reputation) each connection belongs to;
(4) one connection I have probably not noticed yet.
Finish by generating a mind map organised by thread, not by week.
```

**P18 · Teach-back (NotebookLM, Sunday, 10 minutes; replaces part of P4 once a month)**
```
The last source is the transcript of a 3-minute voice memo in which I explain this week's topic without notes.
Grade it against the other sources:
(1) accuracy, 1-5, with every wrong or vague sentence quoted and corrected from a source;
(2) completeness: the two most important ideas I left out;
(3) analogy quality: did my analogy help or mislead, and why;
(4) one question a biologist would ask me after hearing this, and one a staff ML scientist would ask.
Be direct. No encouragement padding.
```

---

## PART 9 · Branding, connection ladder, endorsement

### 9.1 Profiles (week 0, updated at each gate)

LinkedIn headline (week 0):
`Applied AI engineer becoming an AI scientist for biotech | Building open instruments and datasets for fermentation and lab data, in public | 25 yrs leading delivery, ex-Big Four AI transformation lead`

At week 16 replace the middle with: `Author of FermLit (bioRxiv): 2,000 fermentation papers, one validated dataset`. At week 20 add the arXiv paper. At week 32 add "co-author with [lab]".

About: the thesis statement of Part 3.2 in plain words (3 lines), the gauge level with links, one line on the on-prem angle, one line on what you are looking for (collaborators and, from week 27, Loop Audit clients).

GitHub: real photo, one-line bio, `fermlit` and `batchlens` pinned, profile README with the headline numbers and DOIs.

Google Scholar profile at week 16 (preprints count). ORCID at week 0 (free, 5 minutes, needed by bioRxiv and journals).

### 9.2 Content pillars and cadence (same as v4, re-pointed)

| Pillar | Channel | Cadence | Source |
|---|---|---|---|
| Build log with a number | LinkedIn Sunday (+ X copy) | Weekly | The week's build |
| Learning in public | LinkedIn Wednesday note, 10 min | Weekly | Tuesday/Thursday notes: "what I got wrong about Monod this week" |
| Findings | Medium deep dives | Weeks 2, 8, 13, 18, 35 | Gates |
| Science | bioRxiv/arXiv | Weeks 16, 20, 32 | Papers |
| Proof | GitHub + Zenodo + HF | Continuous | Releases |
| Talk | Event | Week 38 | Everything |

Rules unchanged from v4: one number or one admitted failure per post; archetypes, never named weaknesses; zero EY data.

### 9.3 The connection ladder (founders and scientists)

Each rung is one small ask; never skip a rung.

1. **Read** their paper (Track F). Save the outreach line.
2. **React** with substance to one of their posts or papers (one comment that shows you read it).
3. **Send a finding** relevant to them (a FermLit figure about their organism, a batchlens number) with no ask.
4. **Ask 30 minutes** to hear how they do it today.
5. **Run something free** on a sample of their data or literature.
6. **Ask for feedback** on a preprint draft (scientists like being asked).
7. **Ask for endorsement or co-authorship** when there is a concrete result that touches their work.
8. **Offer the Loop Audit** only after rung 6, and only from week 27.

Targets: 5 founders at rung 4 by week 8; 10 by week 16; 2 at rung 7 by week 24; 3 at rung 8 by week 28.

### 9.4 Connection ladder (established companies, funds, scientific-data companies)

- **Funds (GridX first):** send the FermLit preprint at week 16 to your GridX contact with one question: "which three portfolio companies would find this useful?" Funds love giving their startups free expert attention. The portfolio package (v4) comes at week 39, after case studies exist.
- **Established companies (Bioceres, Insud, Biogénesis Bagó, Terragene, Porta...):** find the head of R&D or QA on LinkedIn; the Paper 2 preprint is the message; ask for 30 minutes on "how you monitor batches today". Week 21 onwards.
- **Scientific-data companies (DataHow, Aizon, TetraScience, Uncountable, Benchling, Novasign...):** these read arXiv. Week 19: apply to 5 with the two preprints as the cover letter. Week 32: apply again with three. Employee track and consulting track are not exclusive; a fractional contract with one of them is a valid Half 2 outcome.

### 9.5 arXiv endorsement: three parallel routes (start week 12)

1. **Co-author route:** a founder-scientist or academic who already publishes on arXiv co-authors Paper 2 or Paper 3; co-authors who are registered can submit.
2. **Cited-author route:** email two authors of the papers you cite most (Prompt P13), with the bioRxiv preprint of Paper 1 as proof you are serious. Endorsement asks are normal and usually answered within a week.
3. **Local academia route:** UBA, UNSAM, Instituto Leloir, CONICET groups that publish in q-bio or cs.LG; ask via a founder's introduction, or attend one seminar and ask in person.

Fallback: bioRxiv accepts computational papers without endorsement; Paper 2 goes there at week 20 if arXiv is not yet possible, and is added to arXiv when an endorser appears. The gate is "preprint public", the target is "endorsed".

### 9.6 Templates (reuse v4 section 9 for build log, gap note, video, targeted message, audit one-pager)

Only one template changes. Targeted message (weeks 16+):
```
Hi [NAME],
I read [their paper]; [one specific line showing it].
I just put out a preprint on [FermLit / hybrid soft sensors]: [one number]. Code and data: [DOIs].
[Their organism / process] appears in [N] of the papers in the dataset; I attached the figure.
If it is useful, I would like 30 minutes to hear how you [monitor batches / record trials] today. No pitch.
[NAME]
```

---

## PART 10 · Money (unchanged in substance from v4, moved to Half 2)

Offers, prices and the "sell weeks not savings" rule are in v4 Part 4. Timing in v5: offer built week 27, first invoice gate week 28 (slips to 29 at most), pilot week 37, fund package week 39, retainer or job decision week 40. The employee track runs in parallel from week 19 with no conflict: a job offer at week 25 that passes the culture filter (v4 section 3.3) is a valid way to end the plan early.

---

## PART 11 · Rescue protocols (for the days the plan meets your brain)

| Situation | Protocol |
|---|---|
| Frozen at the start of a block | Prompt P12. Then 10 minutes on the smallest visible thing. If still frozen after 10, walk 15 minutes and retry once. If still frozen, log it and end the block; no guilt, the plan expects it |
| Shiny object (new tool, new paper, new project) | `PARKING.md` in 20 seconds. Saturday sandbox if the week is done |
| Hyperfocus on the current task past the timer | Allowed once per week, max 3 hours, never after 22:30, and the next day's block is cancelled to pay for it |
| Boredom three blocks in a row on a study unit | The unit is not connected to a build. Skip it, note it in `PARKING.md`, move on. Come back only if a paper needs it |
| Imposter moment ("I am a PM pretending to be a scientist") | Open `PROGRESS.md`, read the gauge level and the DOIs. Then read Part 0.1 point 2. A DOI is not an opinion |
| Lost week | Prompt P14 on Sunday. Never two in a row |
| Overwhelm at a gate | Cut to "only one thing". The gate passes on the artifact, not on the polish |
| Nobody answers messages | Expected until week 16. After the preprint, response rates change. Keep the ladder, never skip rungs |
| A founder says "we already have this" | Ask to see it and ask what it misses. That is a rung-4 conversation, which is the goal |
| Extraction accuracy disappointing | Report it. The reporting-completeness finding does not depend on perfect extraction; the paper is about a measured method |

---

## PART 12 · Gates and metrics (the only numbers you track)

| Week | Gate | Artifact that proves it | Gauge |
|---|---|---|---|
| 4 | Public and reproducible | batchlens v0.1 with DOI and headline number | L1-L2 |
| 8 | Technical report | Report on Zenodo; F1 table public; 5 founders messaged | L2 |
| 13 | Dataset public | FermLit v0.5 on HF + Zenodo; Paper 1 skeleton with figures; endorsement route alive | L2+ |
| 16 | Paper 1 | bioRxiv DOI | L3 |
| 20 | Paper 2 | arXiv ID (endorsed) or bioRxiv ID with endorsement dated | L4 (endorsed) |
| 24 | Co-author agreed | Written agreement; validation plan pre-registered in repo | L4 (path) |
| 28 | First invoice | Any amount | Money 1 |
| 32 | Paper 3 | bioRxiv DOI with co-author | L4 (co-authored) |
| 36 | Journal and JOSS | Two journal submissions, one JOSS | L5 (path) |
| 40 | Decision | Retro with income, pipeline, offers | L5 |

Weekly board (Sunday, 30 seconds): blocks done (target 5/7), days with commits (3/7), Sunday log (1), Wednesday note (1, skippable), founder conversations cumulative (5 by w8, 10 by w16, 20 by w28), gauge level, lost weeks in a row (never 2).

Banned: stars, likes, followers, hours studied, videos watched.

---

## PART 13 · Ignore for 40 weeks

- Vara as a project. Its telemetry and judge ideas live in `fermlit/eval` and `batchlens/report`.
- Any wet-lab work by you. Paper 3's lab is the co-author's.
- Protein design, generative biology, AlphaFold-scale anything. Zavia and Nanogrow's territory; you use embeddings, not design.
- Clinical or patient data (archetype E). Half 3, if ever.
- Building a LIMS, an ELN, a SaaS, a marketplace, a dashboard product. eLabFTW exists.
- A fourth instrument before week 40.
- Hourly consulting. Fixed price only.
- Full courses (MIT 6.874 complete, CS336 complete, biostatistics MOOCs, cloud certifications). Part 4 is all that fits.
- Naming a company's weakness in public. Ever.
- A second side project. El Espejo stays at its current pace and does not count.
- Reading the news about AI. Parking lot.

---

## PART 14 · Day 0 (tomorrow), minute by minute

1. (10 min) Create `cockpit` repo. Paste Part 6 and Part 7 into `PROGRESS.md`. Create the other four files, empty except `TODAY.md`: "Create `batchlens` repo with README skeleton from Part 5.2".
2. (20 min) Create `labloop-notes`, `batchlens`, `fermlit` with README skeletons and `CITATION.cff`.
3. (20 min) `uv` environment; `ollama pull qwen2.5:7b-instruct-q4_K_M`; confirm it answers. ORCID account.
4. (10 min) LinkedIn headline and About from Part 9.1.
5. (10 min) Post 0 (Part 6, week 0).
6. (5 min) `LOG.md` first line. Write Monday's `TODAY.md`: "open `batchlens/sim/monod.py`, write the four ODEs for X, S, P, DO with Monod growth; plot one batch."
7. Close everything.

---

## PART 15 · The one thing to remember

At week 40 nobody will ask how many chapters you read. A scientist will ask **what you measured, how you validated it, and who checked it**. A founder will ask **how many days your instrument takes out of their loop**. A company will ask **what you published**. This document exists so that all three answers are a link, a number and a DOI.

---

## APPENDIX V6-A · The NotebookLM calendar at a glance

| Week | Notebook name | Gate extras | Act notebook created |
|---|---|---|---|
| 0 | `LABLOOP-W00 · Cockpit and positioning` | — | — |
| 1 | `LABLOOP-W01 · A tank in 80 lines` | — | — |
| 2 | `LABLOOP-W02 · One hundred real batches` | Data Table | — |
| 3 | `LABLOOP-W03 · Golden batch, with statistics` | — | — |
| 4 | `LABLOOP-W04 · L1 public and reproducible` | Video + Slides | `LABLOOP-ACT-I` |
| 5 | `LABLOOP-W05 · FermLit corpus and schema` | Data Table | — |
| 6 | `LABLOOP-W06 · The human gold standard` | — | — |
| 7 | `LABLOOP-W07 · The extraction agent, measured` | — | — |
| 8 | `LABLOOP-W08 · Technical report gate` | Video + Slides | `LABLOOP-ACT-II` |
| 9 | `LABLOOP-W09 · FermLit scale-up` | Data Table | — |
| 10 | `LABLOOP-W10 · Agreement and analysis` | — | — |
| 11 | `LABLOOP-W11 · Dataset engineering` | — | — |
| 12 | `LABLOOP-W12 · Endorsement and hybrid model` | — | — |
| 13 | `LABLOOP-W13 · FermLit release and Paper 1 outline` | Video + Slides | `LABLOOP-ACT-III` |
| 14 | `LABLOOP-W14 · Paper 1 full draft` | — | — |
| 15 | `LABLOOP-W15 · Paper 1 review and revision` | — | — |
| 16 | `LABLOOP-W16 · Paper 1 preprint gate` | Video + Slides | — |
| 17 | `LABLOOP-W17 · Paper 2 experiments` | Extra Infographic | — |
| 18 | `LABLOOP-W18 · Paper 2 full draft` | — | — |
| 19 | `LABLOOP-W19 · Recruiter portfolio and review` | — | — |
| 20 | `LABLOOP-W20 · Halfway gate` | Video + Slides | `LABLOOP-ACT-IV` |
| 21 | `LABLOOP-W21 · Protein study and partner scope` | — | — |
| 22 | `LABLOOP-W22 · Protein validation and baselines` | — | — |
| 23 | `LABLOOP-W23 · Extremophile candidate ranking` | Extra Infographic | — |
| 24 | `LABLOOP-W24 · Co-author and validation gate` | Video + Slides | `LABLOOP-ACT-V` |
| 25 | `LABLOOP-W25 · trialkit field module` | — | — |
| 26 | `LABLOOP-W26 · trialkit plate module` | — | — |
| 27 | `LABLOOP-W27 · Loop Audit offer and pipeline` | — | — |
| 28 | `LABLOOP-W28 · First invoice gate` | Video | — |
| 29 | `LABLOOP-W29 · Audit delivery 1` (private) | — | — |
| 30 | `LABLOOP-W30 · Audit readout and validation results` | — | — |
| 31 | `LABLOOP-W31 · Paper 3 co-author draft` | — | — |
| 32 | `LABLOOP-W32 · Paper 3 gate and career refresh` | Video + Slides | `LABLOOP-ACT-VI` |
| 33 | `LABLOOP-W33 · labdocs ingestion and traceability` | Extra Infographic | — |
| 34 | `LABLOOP-W34 · labdocs retrieval evaluation` | — | — |
| 35 | `LABLOOP-W35 · labdocs release and local benchmark` | Extra Infographic | `LABLOOP-ACT-VII` |
| 36 | `LABLOOP-W36 · Journal and JOSS submissions` | Video | — |
| 37 | `LABLOOP-W37 · Pilot or job process` (private) | — | — |
| 38 | `LABLOOP-W38 · Scientific talk and proof` | — | — |
| 39 | `LABLOOP-W39 · Fund portfolio offer` | — | — |
| 40 | `LABLOOP-W40 · Final retrospective and career decision` | Video (+ optional Slides) | `LABLOOP-ACT-VIII`, `LABLOOP-40-WEEKS` |

Always-on: `LABLOOP-THESIS` (only papers you cite and your own preprints).

**Privacy rule for notebooks:** client data, partner unpublished results and anything under NDA never go into NotebookLM without written consent. Use anonymised summaries you wrote yourself.

---

## APPENDIX V6-B · Most-used repos, one table (stars checked 2026-09-13)

| Repo | ★ | What it is | Weeks |
|---|---|---|---|
| ollama/ollama | 180.8k | Local LLM runtime | 0, 9, 33, 35 |
| ggml-org/llama.cpp | 128.1k | Local inference engine | 34, 35 |
| huggingface/transformers | 165.3k | Model hub library (ESM-2) | 16, 21 |
| astral-sh/uv | 89.8k | Python environment manager | 0 |
| scikit-learn/scikit-learn | 67.2k | Classical ML, CV, PCA, GPs, metrics | 2, 4, 10, 17, 22 |
| docling-project/docling | 66.3k | Local document conversion | 15, 29, 33, 35 |
| anthropics/claude-cookbooks | 52.7k | Claude recipes: extraction, tool use | 7, 27 |
| jakevdp/PythonDataScienceHandbook | 49.9k | NumPy/pandas/ML notebooks | 1, 2 |
| datalab-to/marker | 39.7k | PDF to Markdown | 33 |
| google/langextract | 38.6k | LLM extraction with source grounding | 5, 7 |
| stanfordnlp/dspy | 38.0k | Prompt programs optimised against metrics (Stanford) | 7 |
| google-research/tuning_playbook | 30.3k | Honest model tuning habits | 4, 18 |
| HumanSignal/label-studio | 28.3k | Annotation tool | 6 |
| karpathy/nn-zero-to-hero | 24.4k | Neural networks from scratch | 8, 9 |
| marimo-team/marimo | 22.7k | Reactive Python notebooks | 20, 27, 32, 37-40 |
| rlabbe/Kalman-and-Bayesian-Filters-in-Python | 19.3k | State estimation intuition | 8 |
| lightgbm-org/LightGBM | 18.8k | Gradient boosting | 8 |
| karpathy/micrograd | 17.5k | Backprop in 100 lines | 8 |
| vibrantlabsai/ragas | 15.7k | RAG evaluation | 34 |
| EleutherAI/lm-evaluation-harness | 14.0k | LLM evaluation discipline | 14 |
| 567-labs/instructor | 13.9k | Structured LLM outputs | 7, 25 |
| statsmodels/statsmodels | 11.6k | Statistics, MixedLM, multiple testing | 3, 14, 25 |
| doccano/doccano | 10.8k | Annotation tool | 6 |
| drivendataorg/cookiecutter-data-science | 10.1k | Project structure | 0, 4 |
| Future-House/paper-qa | 9.2k | QA over scientific PDFs | 14, 31 |
| py-why/dowhy | 8.3k | Causal inference | 22 |
| rtqichen/torchdiffeq | 6.5k | Neural ODEs | 13 |
| quarto-dev/quarto-cli | 6.0k | Reproducible manuscripts | 8, 13-20, 31, 36 |
| biopython/biopython | 5.2k | Sequences, Entrez, UniProt parsing | 11, 21, 23 |
| grobidOrg/grobid | 5.1k | Scientific PDF parsing | 15 |
| argilla-io/argilla | 5.1k | Human feedback on LLM outputs | 6 |
| huggingface/notebooks | 4.6k | Includes the protein language modeling Colab | 13, 16, 17 |
| karpathy/makemore | 4.3k | Character-level language model | 9 |
| facebookresearch/esm | 4.2k (archived) | ESM-2 reference | 16 |
| cornellius-gp/gpytorch | 3.9k | Gaussian processes | 10 |
| meta-pytorch/botorch | 3.6k | Bayesian optimization | 11 |
| sokrypton/ColabFold | 2.9k | Protein structure in Colab | 23 |
| UKGovernmentBEIS/inspect_ai | 2.8k | Evaluation framework | 9, 14, 34 |
| scverse/scanpy | 2.6k | Single-cell analysis | 19 |
| the-turing-way/the-turing-way | 2.2k | Reproducible research handbook | 0, 4, 24 |
| soedinglab/MMseqs2 | 2.1k | Sequence clustering for honest splits | 22 |
| dynamicslab/pysindy | 1.9k | Equation discovery | 12 |
| yangkky/Machine-learning-for-proteins | 1.7k | Curated protein ML reading list | 17 |
| scikit-learn-contrib/MAPIE | 1.6k | Conformal prediction | 16, 22 |
| elabftw/elabftw | 1.4k | Open-source ELN | 33 |
| intro-stat-learning/ISLP_labs | 1.4k | ISLP labs (Stanford authors) | 3, 4, 5 |
| aangelopoulos/conformal-prediction | 1.1k | Conformal notebooks | 16 |
| **Niche but authoritative** (low stars, chosen for authorship) | | | |
| J-SNACKKB/FLIP | 141 | Official FLIP benchmark splits | 17, 21 |
| ievapudz/TemStaPro | 87 | Published thermostability predictor | 21 |
| OATML-Markslab/ProteinGym | 470 | Reference protein fitness benchmark | 17 |
| ACCLAB/DABEST-python | 413 | Estimation plots | 30 |
| erdogant/pca | 354 | PCA with T²/SPE outliers | 3 |
| Pioreactor/pioreactor | 149 | Open-source mini bioreactor | 1, 26 |
| JustusvLiebig/Soft_Sensor_Experiments | 103 | Industrial soft-sensor model zoo | 8 |
| salvadorgarciamunoz/pyphi | 61 | Pharma-practitioner MSPC | 3 |
| elabftw/elabapi-python | 44 | Official eLabFTW API client | 33 |
| StephenGoldie/indpensim-notebook | 29 | IndPenSim summary notebook (verify author) | 2 |
| DoseResponse/drc | 29 | Reference dose-response package (R) | 26 |
| jrcramos/Hybrid-modeling-of-bioreactor-with-LSTM | 24 | Academic hybrid bioreactor model | 12 |

---

## APPENDIX V6-C · Change log v5 → v6

- **Added:** Prologue (mentor letter, eight-act story, seven threads, PM-to-scientist translation table, resource rules, evergreen shelf and channels, NotebookLM protocol, chapter anatomy).
- **Added to every week (0-40):** Welcome with analogy; where the week sits (from / feeds / returns in); goals (this week / next gate / final); Learn it (Core and Deeper); Run it (repos and Colab with verified stars); NotebookLM kit; mentor check.
- **Expanded:** weeks 21-40 from one table row each into the v5 weekly format, gates and outputs unchanged.
- **Changed:** Part 5.3 NotebookLM setup (weekly + act + thesis notebooks instead of three thematic notebooks); Part 8 gains P16-P18; NotebookLM build lives in the last 20 minutes of Saturday's buffer so the Sunday hour is unchanged.
- **Corrected:** Karpathy video title ("spelled-out", not "spelling-out"); repo moves noted (LightGBM → lightgbm-org, BoTorch → meta-pytorch, RAGAS → vibrantlabsai, GROBID → grobidOrg, Anthropic cookbook → claude-cookbooks, evolutionaryscale/esm → Biohub/esm; facebookresearch/esm is archived).
- **Unchanged:** thesis, four papers, reputation gauge, gates (4, 8, 13, 16, 20, 24, 28, 32, 36, 40), cockpit files, rituals, rescue protocols, "ignore for 40 weeks", Day 0.