# LABLOOP v7 roadmap: hybrid, lead-first

Decided 2026-09-29. **This file governs the week sequence from week 5 onward.** Weeks 0–4 are unchanged. Every later week reuses an existing week folder as its *module* (the specs, notebook and cards stay where they are), so nothing is lost; only the order and the scope change.

## Why the plan changed

The v5/v6 plan was built to make a research scientist: three preprints, two journal submissions, a JOSS paper, three products, protein ML, extremophiles, field trials and a fund offer, all in 400 hours. Week 1 showed the real pace: a "30-minute" biology block took five hours. The plan could not finish on time in any order.

The goal is also different. Peter wants a well-paid, coherent role **helping biotech and pharma implement AI**, with a credible personal brand, and without writing papers that do not pay back. That is closer to an AI transformation or delivery lead than to a research scientist.

**Direction: hybrid, weighted towards the lead role.**

- Credibility comes from *working, measured tools* and *real GridX case studies*, not from papers.
- One publication-grade artifact stays: the FermLit technical report and dataset, each with a DOI. No PhD is needed for that, and an evaluated F1 table is exactly the evidence pharma AI leads are asked for.
- Enough biology and math to build solutions for GridX startups and talk credibly with their scientists, learned **just in time** for each build.
- Program management experience from EY is the core asset; the plan finally uses it (discovery, scoping, delivery, readouts).

## Rules that keep it on time

1. **One core source per study slot, 60 minutes maximum.** Everything else is "deeper" and optional.
2. **The understanding check is the stop signal**, not the end of the chapter.
3. **Curiosity goes to [PARKING](./PARKING.md)** and the Saturday sandbox, never into the core time.
4. **One post every two weeks** is the branding cadence, built from that week's real figure.
5. Weeks 18, 26, 30 and 39 are **buffers**. Use them to catch up, not to add scope.

## The 40 weeks

The *module* column names the existing folder whose specs and cards to use. "New" means the spec is written when the week arrives.

### Act I · Bioprocess AI foundation (weeks 0–4, unchanged)

| Week | Focus | Module | Only one thing |
|---|---|---|---|
| 0 | Cockpit and positioning | week-00 | `TODAY.md` filled |
| 1 | Fermentation simulator | week-01 | `monod.py` with the plot |
| 2 | IndPenSim data and baseline | week-02 | tidy parquet of IndPenSim |
| 3 | Statistical batch monitoring | week-03 | T² limits computed |
| 4 | batchlens first release | week-04 | the release with the DOI |

Study for weeks 1–4 is the trimmed core agreed on 2026-09-29: enzymes and Monod → respiration overview → growth curve → fermentation vs respiration; 3Blue1Brown ch. 1–3 → derivatives as rates → dot product, eigenvectors, PCA → cross-validation and leakage.

### Act II · Regulated AI in pharma (weeks 5–10)

The most marketable skill for pharma: AI that a quality department can trust.

| Week | Focus | Module | Only one thing |
|---|---|---|---|
| 5 | Pharma map and regulated-AI literacy: drug development stages, GMP/GxP, ALCOA+ data integrity, GAMP 5, current FDA/EMA AI guidance. **Market check from Argentina:** 10 remote-eligible postings open to LATAM (Wellfound, Get on Board, LinkedIn remote filter), 5 US postings as the ceiling (BioSpace, Levels.fyi), the latest Sysarmy survey as the local floor, and 5 fractional/consulting rate benchmarks | New | `market-check.md` with the three income routes and the ranges they actually show |
| 6 | labdocs ingestion and traceability | week-33 | the tamper-detecting audit log |
| 7 | labdocs retrieval evaluation | week-34 | the 60 labelled Q&A |
| 8 | labdocs release: a batchlens alarm → SOP → deviation draft with human review | week-35 | the benchmark table |
| 9 | GridX case study 1: discovery and scope with one startup | week-27 (offer and discovery parts) | a one-page scope the startup agrees to, naming the money metric (days of cycle time, failed batches, analyst hours) the work will move |
| 10 | GridX case study 1: prototype and readout | week-29/30 (readout parts) | the readout with the money metric measured before and after (or baseline plus a measured prototype estimate), written up with the startup's permission |

### Act III · Measured LLM extraction, the credibility piece (weeks 11–18)

| Week | Focus | Module | Only one thing |
|---|---|---|---|
| 11 | FermLit related work and schema | week-05 | `RELATED.md` |
| 12 | Human gold standard (**30 papers**, down from 50) | week-06 | 30 gold JSONs |
| 13 | Measured extraction agent | week-07 | the F1 table, whatever it says |
| 14 | Scale-up to a few hundred papers + unit normalisation | week-09 + week-11 | normalised dataset export |
| 15 | Agreement check with a second annotator, or error analysis if none | week-10 | kappa or the error table |
| 16 | **Technical report + dataset release, both with DOI** | week-08 + week-13 | the report on Zenodo |
| 17 | Portfolio and positioning: LinkedIn, three 60-second demos, one-page CV per role type | week-19 (portfolio parts) | the portfolio page live |
| 18 | Buffer | — | the earliest carried-over item |

### Act IV · Market: jobs and paid work (weeks 19–30)

| Week | Focus | Module | Only one thing |
|---|---|---|---|
| 19 | Verified target list; first five applications | week-19 (career parts) | five applications sent |
| 20 | Halfway review and reforecast from real evidence | week-20 (review parts) | the reforecast |
| 21 | AI readiness / Loop Audit offer: one-pager with the case studies as proof | week-27 | the one-pager |
| 22 | Outreach: ten messages + GridX introductions | week-27 + week-16 (introduction parts) | ten messages sent |
| 23 | Proposal and **first invoice gate** | week-28 | the proposal sent |
| 24 | Delivery or sales recovery | week-29 | the loop map with cycle-time numbers |
| 25 | Audit readout | week-30 | the readout delivered |
| 26 | Buffer | — | the earliest carried-over item |
| 27 | Interview preparation: leakage, evaluation, GxP, delivery stories | New (uses week-22 career parts) | ten rehearsed answers |
| 28 | GridX case study 2: discovery | New | a one-page scope |
| 29 | GridX case study 2: delivery | New | the readout |
| 30 | Buffer | — | the earliest carried-over item |

### Act V · Proof and decision (weeks 31–40)

| Week | Focus | Module | Only one thing |
|---|---|---|---|
| 31 | Optional depth: hybrid soft sensor on batchlens (the mechanistic + ML idea from old Paper 2), as a report, not a paper | week-12 (hybrid parts) + week-17 | learning curves with seeds |
| 32 | Talk proposal to a meetup or conference | week-38 | the proposal accepted |
| 33 | Talk preparation | week-38 | the rehearsed 20-minute talk |
| 34 | Talk delivered + follow-ups | week-38 | the talk delivered |
| 35 | Optional: JOSS submission for batchlens (peer-reviewed software, no PhD needed) | week-36 (JOSS part) | the JOSS submission |
| 36 | Job process or pilot | week-37 | culture-filter score after the first interview |
| 37 | Job process or pilot | week-37 | the first weekly readout |
| 38 | Portfolio refresh with everything measured so far | week-32 (career parts) | updated portfolio |
| 39 | Buffer | — | the earliest carried-over item |
| 40 | Final retrospective and career decision | week-40 | the decision memo |

## How each stage earns the "from 5 days to 2" sentence

Every stage ends with a number a pharma buyer understands. Public tools show *capability*; only real engagements produce a *client result*, and the two are never mixed in a claim.

| Stage | Money metric | Evidence type |
|---|---|---|
| batchlens (wk 4) | Hours earlier a failing batch is flagged; false alarms per batch | Capability, public simulated/benchmark data |
| labdocs (wk 8) | Minutes to answer a QA question or draft a deviation, manual vs tool, on the benchmark | Capability, synthetic documents |
| GridX case study 1 (wk 10) | The metric agreed in week 9, before and after | **Client result** |
| FermLit (wk 16) | Minutes per paper by hand (timed in week 12) vs the agent, at the measured F1 | Capability, public literature |
| Audit (wk 25) | Cycle-time baseline and estimated days saved per automation candidate | **Client result** |
| GridX case study 2 (wk 29) | The agreed metric, before and after | **Client result** |

By week 30 that is three capability numbers and up to three client results. The portfolio (week 17), the offer (week 21), interview answers (week 27) and the talk (week 34) are built from these numbers only.

## Gates (replace GATES.md targets from week 5 on)

| Week | Gate | Evidence |
|---|---|---|
| 4 | batchlens public | v0.1 public, measured headline, DOI |
| 8 | Regulated-AI demo | labdocs public with its evaluation table |
| 10 | First real case study | GridX readout, written up with permission |
| 16 | Credibility piece | FermLit technical report + dataset, both with DOI |
| 19 | In the market | five real applications sent |
| 23 | First invoice | proposal sent; invoice only for real agreed work |
| 34 | Public voice | one talk delivered |
| 40 | Decision | offer, paid work or a written reforecast, all backed by evidence |

## Parked for after week 40

Not deleted: the folders stay, and any of these can come back if a real opportunity needs it.

- Paper 1 bioRxiv preprint and Paper 2 arXiv, including endorsement hunting (old weeks 12, 14, 15, 16, 18, 20)
- Journal submissions (old week 36)
- Protein ML, extremophile ranking and the partner co-author Paper 3 (old weeks 21–24, 31, 32)
- trialkit field and plate modules (old weeks 25–26)
- Fund portfolio offer (old week 39)

**Revisit condition:** a GridX startup or an employer asks for one of these, or FermLit attracts a scientist who wants to co-author.

## How the rest of the weeks get built

Weeks 1–4 were rewritten on 2026-09-29 in the **mission format**, the template for every later week:

- **Each day is one mission**: an emoji and a short name, a 5-minute **Hook** tied to Peter's own build, one **Do**, and an observable **Done when**. Boring parts are delegated to AI (Peter reviews) or time-boxed.
- **Study days pair learning with the build it explains**, and stop at the understanding check. Leftover time goes back to the build.
- **Saturday**: catch up, then a guilt-free rabbit-hole slot. **Sunday**: "show it", one visible post, and write next Monday's first action.
- **Roadblock rule** (also on every cockpit week page): does *Done when* need it? No → skip and log. Yes → let AI do the boring part. Still stuck → smallest version in 25 minutes.

**Rolling preparation, two weeks ahead.** The later weeks are *not* all rewritten now: the week 5 market check and the GridX case studies will change them, and 36 detailed weeks up front would add overwhelm. Instead:

| When | What gets prepared |
|---|---|
| During week 3 | The cockpit follows v7 order (week N → module folder); week 5 spec and materials written in mission format |
| Every Sunday of week N | Week N+2 in mission format: missions, trimmed core, materials with exact links |
| Week 10 | Checkpoint after the first case study: adjust weeks 11–20 |
| Week 20 | Halfway review: adjust weeks 21–40 from real evidence (market check, applications, GridX) |
| Week 30 | Last reshuffle: choose job, consulting or both for weeks 31–40 |

## What is still pending

- **The cockpit still numbers weeks the old way.** Weeks 0–4 are identical, so nothing is wrong yet. Before week 5, the app needs to follow this order (show v7 week 5 using the new market-check module, week 6 using the week-33 folder, and so on).
- Week 5, 27, 28 and 29 have no spec yet. Write each one when the week is two weeks away.
- The salary target is unverified until the week 5 market check.