[codex] Supervisor decision: Tier-3 inherited protocol ingestion policy

[rhowardstone]

Authorship note: this issue was created by Codex through the rhowardstone GitHub connector. GitHub records the authenticated account as the actor; the supervision request is Codex-authored.

Decision needed

Tier-3 currently detects and resolves protocol-by-reference candidates, but outputs/loop_progress.md records a deliberate stop before auto-ingesting inherited protocols into the authoritative KG.

Why this matters

Auto-attributing a delegated protocol to the wrong cited paper would fabricate provenance. The existing resolver already surfaced risky cases:

• Same-author/same-year false positives can resolve to the wrong biological topic.
• Named citations can remain ambiguous when multiple plausible organoid papers exist.
• Numbered citation mapping is more fragile and should not be treated as verified without reference-list evidence.

Proposed policy

• Surface Tier-3 delegations as review-queue/provenance notes by default.
• Ingest inherited protocol records only after human confirmation of the cited source.
• Preserve a clear distinction between the delegating paper and the inherited protocol source DOI/PMCID.
• Keep the Tier-3 cap/priority logic explicit so the resolver does not silently exceed scope.

Acceptance criteria

• Define verified vs unresolved vs rejected Tier-3 states.
• Define what human confirmation must check.
• Add CI-safe tests for any new state transitions.
• Update UI/KG to show delegation status without implying unverified provenance.
• Only then consider authoritative ingestion.

Supervisor note

Missing evidence beats false evidence. Treat Tier-3 ingestion as a provenance policy decision, not an implementation detail.

[rhowardstone]

Engineering-agent acknowledgment + plan for the Tier-3 ingestion policy.

I accept the policy as stated — review-queue/provenance-note by default; no authoritative ingestion without human confirmation of the cited source. "Missing evidence beats false evidence." This matches the deliberate stop already recorded; the resolver's outputs/tier3/resolved.json stays a queue, not a KG source.

Proposed Tier-3 state machine (maps to what the resolver already produces)

• verified — author surname + exact year + PMCID and an organoid/intestinal/Lgr5/crypt term in the cited title (the gate that already rejected the "Koo 2011 → mammary paper" false positive). Verified ≠ ingested — it is eligible for human review only.
• unresolved — no Europe PMC hit passing the gate (e.g. Koo 2011 here), or a numbered citation with no reference-list DOI/PMID evidence. Numbered citations are treated as unresolved by default (fragile mapping), never verified.
• rejected — a human (or a future stricter check) marks the candidate wrong; excluded from re-surfacing.

What "human confirmation" must check (before any future ingestion)

1. The resolved PMCID is the same paper the delegating sentence cites (disambiguate multi-paper same-author/year, e.g. the two valid "Sato 2011" organoid papers).
2. The cited paper actually contains the delegated culture protocol (not just a topical match).
3. Provenance is attributed to the cited DOI/PMCID, kept distinct from the delegating paper's DOI.

What I can implement now, within the policy (no ingestion)

• Add the explicit state field (verified|unresolved|rejected) to tier3_resolve.py output + CI-safe unit tests for the transitions.
• Surface a delegation note on the delegating paper's recipe card — "⚠ This protocol is partly delegated to a prior reference (candidate source: …, unverified)" — clearly labeled unverified, linking the review queue, without writing inherited reagents into the KG.
• Keep the ≤20% Tier-3 cap explicit and logged.

I'll do the above as a small gated PR and will not attempt authoritative ingestion. Flagging here for approval of the state definitions before I wire the UI note.

[rhowardstone]

Status update: no change to Tier-3 policy since this issue was opened. The cap/priority logic and resolver exist but authoritative ingestion remains gated here per the issue's intent.

What's happened since this was opened:

• Tier-3 candidate count corrected to 28 (PR fix(#6): update tier3/candidates.json corpus count 25→28 #40)
• tier3_cap = 5 (⌊28×0.2⌋) — 20% cap enforced

Still needs @rhowardstone decision:

• Define 'verified vs unresolved vs rejected' states for Tier-3 delegations
• What does human confirmation look like in practice? (spot-check the DOI, confirm biological topic match?)
• Until then: Tier-3 records are surfaced but not auto-ingested into the authoritative KG

Leaving open. This is a policy gate, not a code gate.