humansys/raise/rai-adversarial-review

Purpose

Falsify an analysis before it drives action. A fresh-context reviewer attacks the findings of a research/analysis/design artifact, spot-checks the cited evidence against the actual code/DB/state, and returns a per-claim verdict (holds/weak/wrong) plus what the analysis missed. Single-shot and bounded — cheap in steps, spends inference only where an adversarial lens pays (SOP dev/sops/using-fable-5.md §6; PAT-E-9552). Not a research loop: it reviews a result, it does not re-derive it.

Mastery Levels (ShuHaRi)

• Shu: Walk each claim; explain each verdict and the spot-check that grounds it.
• Ha: Batch obvious holds; expand only weak/wrong and the blind-spot.
• Ri: Verdict table + blind-spot + overreach, minimal prose.

Context

When to use: After a research/analysis/design phase (spike research, architecture proposal, audit report) — before committing to its conclusions.

When to skip: Reviewing code changes (use rai-quality-review) or design proportionality (use rai-architecture-review). This skill reviews conclusions, not diffs.

Inputs: artifact (path to the analysis/design doc). Optional: rubric (what to attack hardest) and sources (paths/DB the claims cite, so verification is cheap).

Steps

Step 0: Portfolio context — fail-open (si $JIRA_KEY disponible)

rai portfolio suggest "$JIRA_KEY" 2>/dev/null || true

Si produce output → usar components_touched para verificar si los módulos citados en el artifact coinciden con el perfil de portfolio (desalineación = candidato a weak/wrong). Usar change_mode para calibrar intensidad: breaking → escrutinio alto en contratos públicos; evolutionary → contexto normal. Si $JIRA_KEY no está disponible o no produce output → continuar sin contexto de portafolio — no bloquear.

Step 1: Load target + extract claims

Read the artifact. Enumerate its load-bearing claims/findings (the ones action depends on). Ignore prose; target assertions with consequences.

Step 2: Attack each claim (default to skepticism)

For every load-bearing claim, try to refute it. Where it cites evidence (a path+symbol, a DB row, a count, an ADR), verify that citation yourself with Read/Grep/Bash before agreeing — reproduce the number, open the file at the line, run the query. Only mark holds when the evidence checks out against reality. Do not be agreeable; a plausible-but-unverified claim is weak, not holds.

# Example spot-checks (adapt to the artifact's citations)
grep -n "<cited symbol>" <cited path>
sqlite3 ~/.rai/raise.db "SELECT count(*) FROM <cited table> WHERE <cited predicate>"

Step 3: Hunt the blind-spot and overreach

• Missing: what did every finding assume in common (groupthink), or which modality/claim

went unverified? Name it.

• Overreach: does a recommendation discard a real constraint for elegance? Name the

baby-with-bathwater risk.

Step 4: Return the verdict (schema, JSON only)

Return ONLY the JSON below — no prose, no preamble. Provide justification only in the critique/evidence/fix fields (typed content). Never reproduce, echo, or narrate your internal reasoning as response text — that is content vs. transcription, and transcription requests trigger a refusal/fallback (SOP §4).

{
  "confidence": "one line: how trustworthy is the artifact, with the biggest remaining unknown",
  "claims": [
    {"target": "claim by title/id", "verdict": "holds|weak|wrong",
     "critique": "the attack", "evidence": "what you checked and found", "fix": "how to correct/strengthen"}
  ],
  "blind_spot": "what all findings share as an unstated assumption, or what went unverified",
  "overreach": "a real constraint discarded for elegance, or 'none'"
}

Output

Persistence of multi-model adversarial reviews

When running adversarial reviews across multiple models (e.g. Fable + Kimi K3 + Codex Sol), persist all artifacts under the epic that invoked them:

work/epics/e{N}-{slug}/
  └── adversarial-reviews/
      └── <topic>/
          ├── prompt.md          # The review brief sent to all models
          ├── fable.md           # Fable review
          ├── kimi-k3.md         # Kimi K3 review
          ├── codex-sol.md       # Codex Sol review
          └── synthesis.md       # Optional: consolidated findings

These are decision-support artifacts — they document the adversarial reasoning behind product and design choices. Persist them in the work context that invoked them, not in the session scratchpad or in dev/docs/.

Quality Checklist

• [ ] Every load-bearing claim has a verdict grounded in a spot-check, not deference
• [ ] Cited numbers/paths reproduced against real code/DB before any holds
• [ ] Blind-spot and overreach both stated (or explicitly "none")
• [ ] Output is JSON only; no "show your reasoning" instruction anywhere
• [ ] Reviewed the result, did not re-run the research

References

• SOP: dev/sops/using-fable-5.md (§4 gotchas, §6 adversarial framing)
• Pattern: PAT-E-9552 (gather cheap → one Fable shot)
• Sibling reviews: rai-architecture-review (proportionality), rai-quality-review (diffs)