Labdocs retrieval evaluation

This week delivers the three concrete artifacts below and the supporting learning/branding/opportunity pack. The scientific or commercial outcome follows the preserved source brief, while the acceptance checks make its evidence testable. The career objective is to turn this work into something a scientist, hiring manager or buyer can inspect.

• Scientific/build priority: Prepare 60 human-labeled questions across document types, including missing-answer, contradictory, versioned and adversarial-content cases.
• Public/profile priority: Draft a citations-versus-correctness post and one failed retrieval case.
• Career/relationship/commercial priority: Prepare a domain review of the evaluation rubric; map this artifact to relevant employer responsibilities.

Read v6 · Week 34 · labdocs, part 2 before Monday. It expands this week's one-row v5 brief into daily build steps and adds the Welcome analogy, Learn it (▶ core and ＋ deeper sources), Run it (repos with a 20-40 minute exercise), the NotebookLM kit and the mentor check below. Standing v6 rules: resource rules, NotebookLM protocol, chapter anatomy. v6 adds depth only; it does not change a goal, gate, output or the 600-minute cap, so this specification still governs execution.

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

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.

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 33: inspect its actual evidence and review; source/code artifacts must exist before the dependent task begins.
• Read commitment, evidence, review and the current root TODAY/PROGRESS/QUESTIONS files.
• Read this week's chapter, v6 · Week 34 · labdocs, part 2, then the standing sections: Part 0 · how to use the document, Part 3 · the scientific thesis, Part 4 · curriculum, Part 8 · the prompt pack and Part 9 · branding and endorsement. The same sections exist in the v5 master, which remains the governing sequence; v6 reproduces them and adds depth. Retained commercial/instrument details come from v4.
• Inspect the actual canonical repository state, available data, existing figures and required human/external inputs. Never assume a previous week succeeded from its folder name.
• Verify current resource links, versions, dataset rights and any live company/venue information before relying on them. Use source catalog; explicitly record inaccessible resources.

Biology/domain unit

Material: Study information retrieval evaluation and Stanford evaluation materials by topic, not assumed lecture numbering; compute recall@k and support labels on five examples.

Produce: One concise applied note, a worked decision/example and unresolved questions.

Understanding check: The note directly supports this week's experiment, collaboration, delivery or career action.

During preparation, write one worked example connected to the week's actual task, one exercise with a separate answer key and five questions. Peter supplies answers; score explanation quality and flag misconceptions in QUESTIONS. Numeric/scientific conclusions require actual data or explicitly synthetic teaching values.

AI/scientific-method unit

Material: Calibrate the judge against the human rubric and inspect disagreement; last 30 minutes: relevant scientific document-workflow paper.

Produce: One concise applied note, a worked decision/example and unresolved questions.

Understanding check: The note directly supports this week's experiment, collaboration, delivery or career action.

Use the prompt pack: P1 explanation ladder, P2 concept-to-project bridge and P3 quiz. Coding practice belongs within the allocated study/build block. Where the source asks for a favorable scientific result, substitute an honest measured comparison as explained below.

Founder/scientist paper

Select one verified founder/scientific-lead paper relevant to this week's domain. Use P11; save the question, method, numerical result with context, data workflow, three author questions and one connection sentence in founder-paper.md. Keep abstract-only access limitations explicit. On source weeks with no study, prioritize the commercial/submission task and log any skipped digest.

Branding: Draft a citations-versus-correctness post and one failed retrieval case.

Opportunity action: Prepare a domain review of the evaluation rubric; map this artifact to relevant employer responsibilities.

Use career asset contracts. Publish at most the agreed cadence; one figure and one supported number or admitted failure can support multiple drafts. In long-form weeks, prepare a 1,500–2,500-word article from existing evidence within the build/buffer allocation. In release/demo weeks, prepare a 60-second script: problem, input, command/action, result, limitation and proof link.

Verify individuals, jobs and affiliations at execution time. The GridX list and v4 company map are research seeds, not current facts. Do not reveal named company weaknesses or employer/client confidential information. Keep drafted/sent/replied/booked/held statuses separate. The planning request alone does not authorize external messages or publication; honor explicit authorization in the execution session.

The 60 cases are a small measured benchmark, not universal accuracy. Keep tuning cases out of final scores and do not treat judge agreement as factual correctness.

Apply the root scientific evidence contract: provenance; meaningful independent splits; no future-data leakage; baseline comparison; actual denominators/uncertainty; failed cases; human review where required. State whether data are simulated, public experimental, synthetic or partner-provided.

Keep source gates intact, and log scope/time conflicts in review. Missing people, money, access or results must not be replaced with invented evidence. Reduce optional polish first. A specific essential external dependency stays pending while independent preparation continues.

This is an enabling week. Its deliverables feed the next scheduled gate: 36. Complete the acceptance checklist below; do not promote reputation from a draft.

• Monday artifact satisfies: Produce the human-checked retrieval score table for the cases actually completed and log any shortfall against the 60-case target.
• Wednesday artifact satisfies: Retrieval recall, citation support and answer correctness have separate counts.
• Friday artifact satisfies: Model/prompt versions and held-out agreement are reported with failure examples.
• Study pack is prepared; Peter's actual learning/quiz result is recorded separately.
• Branding and opportunity drafts use supported claims and the correct external-action state.
• Reproduction or document checks are recorded with file/version references.
• This week's special gate, if any, has real evidence or remains explicitly pending.
• Review, root progress/log and the next physical action are updated.

Expanded minimum derived from the source goal: Produce the human-checked retrieval score table for the cases actually completed and log any shortfall against the 60-case target.

Operational first step if stalled: open retrieval-gold-set.csv, write the required inputs and produce the smallest checkable example in ten minutes. If a more urgent missing gate dependency controls the week, use the source minimum above.

If the full week does not fit, use P14: retain one gate-critical deliverable and one directly supporting study exercise; record what is skipped versus moved. Do not mark the original full gate passed merely because the minimum was completed. After a missed week, reforecast from actual evidence and avoid a second unexamined lost week.

In review, state exactly what exists, what was verified, what Peter can now explain, which external events remain pending, where the canonical code/data live, and the next file/function or draft paragraph to touch. Link each fact to evidence. Carry scientific assumptions, split definitions and rights constraints forward explicitly.

Prepare Week 34 of this LABLOOP execution plan.
Read execution-plan/specs.md, execution-plan/DECISIONS.md,
execution-plan/GATES.md and
execution-plan/weeks/week-34-labdocs-retrieval-evaluation/specs.md, commitment.md, evidence.md and review.md.
Read prior prerequisite reviews and inspect actual repository/data state.
Use v5 as the governing sequence, the v6 Week 34 chapter
(LABLOOP-Master-Plan-v6.md) as the deeper walkthrough, and v4 only for
retained details.
Produce the full local preparation pack and the exact week-specific files named
in the artifact table. Give me a bounded lesson with exercises/answer key,
daily physical actions, technical runbook or draft, branding drafts and
career/commercial actions. Keep the total to 600 minutes.
Separate prepared artifacts from measured results and my learning completion.
Verify sources and do not invent metrics, citations, papers, people or outcomes.
Do not execute external communications, publication or spending unless the
current session already authorizes the specific action.
Finish by updating evidence/review and TODAY with the next physical action.

v6 expands this row into a full brief, with the gate and output unchanged, inside v6 · Week 34 · labdocs, part 2. The following is source planning text, not a claim of completion. Apply the explicit evidence corrections in this specification and DECISIONS.md.