sdrf-adversarial-review

name

sdrf-adversarial-review

description

Independently falsify and review a created, modified, or proposed SDRF against its specification, ontology terms, source evidence, repository files, and experimental design. Use after SDRF annotation or repair, before contribution or completion, when a review gate reports a pending artifact, or whenever an isolated reviewer must distrust the producer's self-assessment and issue a hash-bound verdict.

SDRF Adversarial Review

Review the artifact from fresh context. Try to disprove its correctness; do not polish or extend it. Approve only what the artifact and cited evidence support.

Preserve Independence

1. Confirm that this run is a fresh reviewer or hook-agent context.
2. Use only the original task, SDRF, specification, evidence manifest, repository files, and deterministic outputs supplied to the reviewer.
3. Do not read or rely on the producer's transcript, reasoning, verdict, or summary. Treat producer claims as untrusted.
4. If context isolation cannot be established, return REVIEW_UNAVAILABLE and do not create a passing receipt.
5. Do not edit the SDRF. Return findings to the producer.

Read references/review-contract.md before writing the report or recording approval.

Review Workflow

1. Freeze the artifact

Compute its SHA-256 and record the repository-relative path. If the hash changes during review, discard the review and start again.

2. Reconstruct requirements

• Read spec/sdrf-proteomics/TERMS.tsv.
• Read spec/sdrf-proteomics/sdrf-templates/templates.yaml and every active template YAML.
• Derive active templates independently; do not accept the producer's list without checking the file.
• Read the original request and evidence manifest, if supplied.

3. Run deterministic checks

Run official parse_sdrf validate-sdrf for every active template. Also run:

python3 -m tools check <artifact>
python3 -m tools score <artifact>

If the repository tools are not importable from the target project, invoke them from the loaded plugin:

PYTHONPATH="$CLAUDE_PLUGIN_ROOT" python3 -m tools check <artifact>
PYTHONPATH="$CLAUDE_PLUGIN_ROOT" python3 -m tools score <artifact>

Do not mark deterministic_validation as passing when required validation could not run.

4. Perform six falsification passes

1. Specification compliance — missing or extra columns, template/version errors, reserved words, structured-value syntax, row shape, and duplicates.
2. Ontology integrity — resolve every unique controlled value and confirm label, accession, ontology family, obsolescence, and justified specificity.
3. Source evidence — challenge every nontrivial biological and technical claim against PRIDE, publication, supplementary data, or raw-file evidence. Flag per-sample values inferred only from cohort summaries.
4. File mapping — compare every comment[data file] with repository files; look for missing, duplicate, unmatched, or non-raw files.
5. Design consistency — recompute sample counts, factors, biological and technical replication, fractions, labels, and potential confounders.
6. Omission safety — search for unsupported additions disguised as specificity and for material evidence the producer omitted.

For each finding, give a precise row/column or file-level location, evidence, and a correction criterion. Classify it as:

• blocker: invalid, unsafe, fabricated, or structurally unusable.
• important: required before contribution or completion.
• minor: useful improvement that does not block approval.

5. Decide and record

Return FAIL if any blocker or important finding remains, or if specification compliance, ontology integrity, omission safety, or deterministic validation is unverified. Do not record a receipt. Give the producer an actionable finding list.

Return PASS only when all required checks in the review contract pass and no blocker or important finding remains. Write the report JSON, then bind approval to the current artifact hash:

python3 "$CLAUDE_PLUGIN_ROOT/tools/review_gate.py" approve <artifact> \
  --report <report.json> --reviewer independent-hook-agent --cwd <repo-root>

Outside Claude Code, replace $CLAUDE_PLUGIN_ROOT with the installed sdrf-skills root. A producer agent must never run approve for its own work.

After any producer edit, repeat this skill in a new reviewer context; the old receipt is intentionally stale.