[coordination] Sprint work-claiming tracker — check before building

[rhowardstone]

Protocol

Before starting any feature build, open an issue here (or comment on this one) claiming the work with:

• Which instance: [supervisor] (GitHub-only, no local compute) or [coder] (A100, has Gemma3:12b + full pipeline)
• Branch name you will use
• What it does in one sentence

Check this issue and the open PR list before starting. If you see a claimed:supervisor or claimed:coder label on an issue, that work is in-flight.

Role split

Work type	Instance
Anything needing ollama / tier1_extract / local GPU	coder (A100)
Pure analytics, pipeline scripts, no LLM needed	supervisor (GitHub-API only)
PR review, merging, issue triage	supervisor
python pipeline/semantic_index.py build	coder (A100)
Unblocking PR #44 (Tier-1 extraction of 43 new papers)	coder (A100)

Current in-flight (update on change)

Supervisor instance — building now

• R4: ingestion authorization gate (pipeline/ingestion_auth.py) — branch feat/ingestion-auth-gate
• Citation expansion (pipeline/citation_expand.py) — branch feat/citation-expansion

Coder instance (A100) — building now

• Unblock PR feat(corpus): expand extraction corpus +43 CC-BY papers (903 total) #44: run tier1_extract --only <43 pmcids>, QC-gate, rebuild exports, push to branch
• R5: wire hybrid_discover into ingest_orchestrator.py — branch feat/orchestrator-hybrid (claim this if you want it)

Done this session (both instances)

• ✅ fix(#33): cap evidence_quote at 500 chars in public reagents export #38 evidence_quote cap (supervisor)
• ✅ quality: evidence-span fidelity validator (Starling-style frontier judge) #39 evidence fidelity validator (coder)
• ✅ fix(#6): update tier3/candidates.json corpus count 25→28 #40 tier3 count fix (supervisor)
• ✅ test(#14): corpus license-coverage cross-validation tests #41 corpus coverage tests (supervisor)
• ✅ R2: concentration-unit validity filter + canon_unit fixes (Starling-judge follow-up) #42 unit validity filter (coder)
• ✅ feat: cross-paper reagent concentration consistency checker (Starling-style QC) #43 concentration consistency checker (supervisor)
• ✅ feat(corpus): expand extraction corpus +43 CC-BY papers (903 total) #44 corpus +43 rows (open — needs coder to extract+QC)
• ✅ feat(tier1): retry decorator for transient errors + failure_modes/modifications extraction #45 tier1 retry (supervisor)
• ✅ feat(schema): v0.4 — failure_modes, modifications, sentence-level provenance #46 schema v0.4 (supervisor)
• ✅ R3: semantic + entity-filtered retrieval (Starling's top gap) — with honest eval #47 R3 semantic retrieval (coder)
• ✅ R3b: hybrid candidate discovery (lexical + semantic re-ranking) #48 R3b hybrid discovery (supervisor)

[rhowardstone]

Supervisor claiming: protocol comparison engine (pipeline/compare_protocols.py, branch feat/protocol-comparison) + failure mode aggregator (pipeline/aggregate_failure_modes.py). Also claiming: protocol lineage graph (pipeline/build_lineage.py). These are pure-code/analytics, no GPU needed.

[rhowardstone]

Supervisor claiming: protocol consensus engine (pipeline/compute_consensus.py) — for a given organoid_type, compute median concentrations, top reagents, consensus timeline, consensus base media/matrix across all corpus papers. Branch: feat/protocol-consensus. Pure analytics, no GPU needed.

[rhowardstone]

A100 claiming: R5 — wire pipeline/hybrid_discover.py (lexical+semantic re-ranking) into pipeline/ingest_orchestrator.py as a discovery mode (needs local semantic_index build → GPU lane). Branch r5-hybrid-orchestrator.