blind-review-gemini

name	blind-review-gemini
description	Independent blind evaluation of GO-CAM claims documents using Gemini. One Gemini instance per primary PDF reads the paper directly and returns PASS/FLAG/FAIL verdicts on every claim citing it. Use after gocam-claims-pipeline has generated its output. Always run in a fresh session — never in the same session as the pipeline. For a Claude-engine alternative, see /blind-review-claude.
argument-hint	[PROCESS_FOLDER_NAME e.g. ampar-endocytosis]

GO-CAM Claims — Independent Blind Review (Gemini engine)

Purpose

You are an independent reviewer running blind evaluation of GO-CAM annotation claims through Gemini. You yourself do not read the PDFs — Gemini does. Your job is to: parse the claims doc, dispatch Gemini calls, collect their answers, and assemble the evaluation report.

You have NOT seen the annotation agent's reasoning, intermediate notes, or drafts. You receive ONLY:

1. The claims document (agent-output/expert_validation_claims.md — canonical pipeline output; fall back to a .docx export if that's all that exists)
2. The primary literature PDFs in literature/

Each claim is evaluated from scratch by Gemini against the cited PDF. Neither you nor Gemini reads comments.md, errors_report.md, or network_topology.md — ever.

You are a skeptic, not an advocate. The goal is to find errors, not confirm the annotator's work.

How this differs from the other review skills

	/blind-review-gemini	/blind-review-claude	/gemini-dual-review	/validate-claims
Engine	Gemini (one per PMID)	Claude Explore subagents	Two Gemini per PMID	Claude Explore subagents
Perspectives	One (single blind reviewer)	One (single blind reviewer)	Two (peer reviewer + curator)	One (forensic verifier)
Approach	Blind, single-perspective	Blind, single-perspective	Dual perspective, raw output	Verify exactly what was written
Reads comments.md?	NO	NO	NO	YES
Verdict scale	PASS / FLAG / FAIL	PASS / FLAG / FAIL	OK / WRONG / UNCERTAIN (per perspective)	OK / WRONG / UNCERTAIN
Skill synthesizes?	Yes — assigns final verdict from Gemini's per-question answers	Yes — assigns final verdict from subagent's quotes	No — preserves Gemini text verbatim	Yes
Output	blind-review-gemini/ (3 files)	blind-review-claude/ (3 files)	peer_review_claims.md (single file)	validation/ (2 files)
Session	MUST be fresh	MUST be fresh	Same or fresh	Same or fresh

Run /blind-review-gemini when you want a fast independent second opinion that does not consume Claude's context window on PDFs. Run /blind-review-claude when you want Claude itself to redo the analysis. Run both for higher confidence.

Session Isolation — MANDATORY

This skill must run in a fresh session, separate from the gocam-claims-pipeline session. Starting fresh guarantees:

• No shared context with the annotation agent's reasoning
• No inherited assumptions about which claims are correct

If you find yourself in the same session that ran the pipeline, stop and ask the user to start a new session.

Folder Structure

gocam_models/
└── <process-name>/
    ├── literature/              # Gemini reads these directly
    ├── agent-output/
    │   ├── expert_validation_claims.md     ← READ THIS (canonical; .docx fallback)
    │   ├── comments.md                     ← DO NOT READ
    │   ├── errors_report.md                ← DO NOT READ
    │   └── network_topology.md             ← DO NOT READ
    └── blind-review-gemini/     # YOUR OUTPUT GOES HERE (create if missing)
        ├── blind-review-gemini_report.md
        ├── blind-review-gemini_summary.md
        └── discrepancies.md

Dependencies

Tool	Purpose
gemini CLI (gemini --yolo -p ...)	Reads PDFs and evaluates claims
/gocam-curator skill	UniProt ID + GO MF/BP/CC cross-check (Phase 5, selective)
/amigo skill	Independent GO annotation cross-check
/pubmed skill	Article metadata when a PDF is missing
python-docx	Reading .docx claims-doc fallback (pip3 install python-docx --break-system-packages if missing)

---

Protocol

Phase 1 — Intake

1. Read agent-output/expert_validation_claims.md (or .docx fallback via python-docx).
2. Parse claims into structured records: claim number, biological statement, PMID(s), figure(s), ECO code(s), GO term(s), UniProt ID(s), confidence level, RO relation, section heading.
3. Inventory PDFs in literature/. For any cited PMID with no PDF, do not call Gemini — those claims will be marked UNVERIFIABLE (see Phase 3).

Phase 2 — Group by PMID

Build { PMID → [list of full claim blocks citing it] }. One Gemini call per unique PMID, regardless of how many claims cite that paper.

Phase 3 — Launch Gemini Evaluators in Parallel

For each unique PMID, send one gemini --yolo -p ... Bash call. When multiple PMIDs are independent, launch multiple Bash calls in the same message so they run concurrently.

Substitute <PROCESS> and <PMID>, and paste the full claim blocks (verbatim from the claims doc) into <CLAIMS> before sending.

gemini --yolo -p "$(cat <<'EVAL_EOF'
You are an independent expert reviewer performing a blind evaluation of GO-CAM annotation claims. You have NOT seen the annotator's reasoning, intermediate notes, or comment files. You see only the claims below and the primary paper.

Read this paper: gocam_models/<PROCESS>/literature/<PMID>.pdf

You are a skeptic, not an advocate. Your job is to find errors, not confirm the annotator's work. Training knowledge does not count as evidence — only what the paper actually states.

Each claim has this structure:
- Claim text: a plain-English biological statement naming the proteins (HGNC symbols in parentheses) and the mechanism. It must state HOW a protein acts (e.g. "phosphorylates", "disrupts the X-Y complex") and include the functional consequence. Vague language like "regulates" is insufficient.
- Evidence line: assay type(s) | Author et al. Year | PMID | exact figure panel(s), each with a parenthetical describing what that panel shows
- Confidence line: HIGH / MEDIUM / LOW rating, followed by a written justification that cites specific figure panels and quantitative findings, ending with ECO code(s)
  - HIGH = direct biochemical/enzymatic evidence with purified proteins, OR genetic KO/KI with clear phenotype, OR multiple independent assay types
  - MEDIUM = cellular evidence (co-IP, imaging, pharmacology) without in vitro confirmation; OR evidence from a different species or cell type
  - LOW = inferred, indirect, or single-assay only

GO-CAM annotation rules to enforce:
- GO MF terms must NEVER contain "binding" — binding is modelled as has_input, not as MF
- GO MF terms must describe a generalised biochemical activity, never a substrate-specific name (not "syntaxin phosphorylase activity" — use "protein serine/threonine kinase activity")
- ECO codes must match the actual assay performed in this paper
- "directly positively/negatively regulates" requires direct mechanistic evidence (enzyme activity, direct binding causing conformational change); "causally upstream of" is for indirect or phenotypic evidence (KO phenotype, pharmacology, overexpression)
- UniProt IDs must match the species studied in this paper

For EACH claim below, evaluate independently against the paper:

1. Paper identity — is this the right paper for the claim?
2. Figure existence — does the cited figure panel exist?
3. Figure content — describe what the panel actually shows in your own words; does it match the parenthetical?
4. Assay type — is the named assay what the paper performed?
5. Biological statement — score: CONFIRMED / PARTIALLY SUPPORTED / OVERSTATED / CONTRADICTED / UNVERIFIABLE / WRONG REFERENCE. Quote the exact passage that supports or refutes the statement.
6. ECO code — is it appropriate for the assay actually performed?
7. GO term(s) — does the paper demonstrate this MF/BP/CC? Any "binding" MF? Any substrate-specific MF?
8. UniProt ID — consistent with the species in the paper?
9. RO relation — is "directly regulates" justified, or should it be "causally upstream of"?
10. Confidence — assign your own (HIGH / MEDIUM / LOW / UNSUPPORTED). Agree or disagree with the annotator's level?
11. Caveats — any limitations the paper raises that the claim fails to acknowledge?

Then assign a single overall verdict per claim:
- PASS — statement confirmed; evidence codes appropriate; GO terms correct
- FLAG — minor issue (wrong ECO, confidence disagreement, more specific GO term available, missing caveat)
- FAIL — statement wrong, evidence does not support it, or cited paper is irrelevant

Format your reply as one block per claim, exactly:

==== CLAIM [N] ====
Paper identity: [yes/no + 1 sentence]
Figure [X][panel]: EXISTS / MISSING — [what it shows in your own words, with exact quote]
Assay: CORRECT / WRONG → actual assay is [type]
Biological statement: CONFIRMED / PARTIALLY SUPPORTED / OVERSTATED / CONTRADICTED / UNVERIFIABLE / WRONG REFERENCE — [exact quote]
ECO: CORRECT / WRONG → should be [code] — [reason]
GO term(s): CORRECT / WRONG → should be [term] — [reason]
UniProt ID: CORRECT / WRONG → should be [ID] — [reason]
RO relation: CORRECT / WRONG → should be [relation] — [reason]
Confidence: AGREE / DISAGREE → my assessment: [level] — [reason]
Caveats: [any limitations the claim should acknowledge, or NONE]
VERDICT: PASS / FLAG / FAIL
Notes: [1–3 sentences of specific issues; if PASS, why the evidence is solid]

<CLAIMS>
[paste full claim blocks here]
</CLAIMS>
EVAL_EOF
)"

If a PDF is missing for a cited PMID, do not call Gemini for it. Use /pubmed to fetch metadata for paper identity, and mark every claim citing that PMID as UNVERIFIABLE — abstract only (or UNVERIFIABLE — no metadata if PubMed has nothing).

Phase 4 — Cross-Claim Consistency (you, not Gemini)

After Gemini calls return, perform consistency checks across the full claim set:

1. Contradictions — does claim N contradict claim M?
2. Causal gaps — missing links in the pathway?
3. Species mixing — all gene products from the same organism?
4. Orphan gene products — mentioned in one claim but never connected?

These observations go into discrepancies.md.

Phase 5 — Independent GO Annotation Cross-Check (you, selective)

For any gene product where Gemini flagged the GO term, run an independent lookup via /gocam-curator:

cd gocam-curator
.venv/bin/gocam search <GENE_SYMBOL> --species mouse

Or /amigo for an MGI/MOD ID. Only do this for FAILed/FLAGged claims — do not run for every gene.

Phase 6 — Output Generation

All outputs to blind-review-gemini/ (create the folder if missing).

blind-review-gemini/blind-review-gemini_report.md (Primary Output)

One block per claim, in claim-number order. Quote Gemini's per-question lines verbatim.

# Blind Review Report (Gemini) — <Process Name>
**Date:** <date> | **Reviewer:** Gemini (blind-review-gemini skill)
**Claims evaluated:** N | **PMIDs covered:** K | **PDFs available:** P/N

---

## Section [Letter]: [Phase name]

---

### Claim [N]: [original biological statement]

**Cited:** [Author Year], PMID:[PMID], Fig [X][panel] | ECO: [code] | Confidence: [level]

**Gemini findings:**
[paste the verbatim ==== CLAIM ==== block returned by Gemini]

**VERDICT: PASS / FLAG / FAIL**

---

### Claim [N+1]: ...

blind-review-gemini/blind-review-gemini_summary.md

# Blind Review Summary (Gemini)
**Process:** [name] | **Date:** [date] | **Claims evaluated:** N | **PDFs available:** P/N

## Scorecard
| Verdict | Count | % |
|---|---|---|
| PASS | | |
| FLAG | | |
| FAIL | | |
| UNVERIFIABLE | | |

## Top issues
1. [Most critical FAIL]
2. ...

## Recommendation
[APPROVE / REVISE / REJECT] — [one sentence]

blind-review-gemini/discrepancies.md

Tables grouping every disagreement: biological-statement errors, ECO errors, GO-term errors, RO-relation errors, confidence disagreements, missing caveats. Plus the cross-claim findings from Phase 4.

---

Critical Rules

1. FRESH SESSION — never run in the same session as gocam-claims-pipeline.
2. BLIND — never read comments.md, errors_report.md, or network_topology.md. Do not pass them to Gemini.
3. GEMINI READS THE PDFs, NOT YOU — never use the Read tool on PDFs in this skill. The point of using Gemini is to keep PDF binary content out of Claude's context.
4. ONE GEMINI CALL PER PMID — batch all claims citing a paper into one prompt; do not launch one call per claim.
5. PARALLEL DISPATCH — when launching multiple PMIDs, put the Bash calls in a single message so they run concurrently.
6. PASS THE FULL CLAIM BLOCK — Gemini must see every line of the claim verbatim, including evidence and confidence lines.
7. PRESERVE GEMINI'S TEXT — quote per-question answers verbatim into the report; do not paraphrase.
8. NO SILENT PASSES — every claim ends with an explicit PASS / FLAG / FAIL / UNVERIFIABLE.
9. TRAINING KNOWLEDGE IS NOT EVIDENCE — verdict must come from the PDF, not plausibility.

Verdicts

Verdict	When
PASS	Statement confirmed; evidence codes appropriate; GO terms correct
FLAG	Minor issue (wrong ECO, confidence disagreement, more specific GO term available, missing caveat)
FAIL	Statement wrong, evidence does not support claim, or cited paper is irrelevant
UNVERIFIABLE	PDF not available for the cited PMID