Decision Review (reliability discipline)
Makes design decisions reliable, not just stated. Two parts: a per-decision 6-field floor (checkable),
and a cross-model red-team critique before publish. Self-built, borrowing the ideas of structured-MADR
(machine-checkable decision fields) and adversarial-review (cross-model critique) — see the research catalog borrow-note.
Pairs with structuring-solution-docs (its S8/MADR blocks ARE these decisions). Generic (skill_spec §9).
When to use
- The research-solution agent finalizes the
solution_document(Wildtarsier §1.9); apply before 准出 (Gate 2).
A. Per-decision 6-field floor (准出 completeness — missing any = not done)
Every key design decision MUST carry all six. The 准出 check (Gate 2, compliance) verifies presence — this is a
completeness criterion of the solution_document output, NOT a new gate (the loop stays the existing two gates):
- Options compared — ≥2 candidates with pros + cons each (no single-option "decision").
- Evidence with provenance — each claim sourced + tagged 实测 / 文献 / 推测 (defer to technical-report-style R4).
- Trade-offs — what the chosen option costs, stated plainly (no one-sided upside).
- Assumptions — what must hold for the decision to be valid (so they can be checked later).
- Failure modes — how it can go wrong + the mitigation or detection.
- Validation — how the decision will be verified (experiment / metric / PoC) — ties to the feasibility region.
B. Cross-model red-team (pre-publish adversarial pass)
- Before the solution_document is published, a different agent / model reviews it adversarially: actively look for weak evidence, unstated assumptions, missing failure modes, one-sided trade-offs, and over-claims.
- Findings are recorded; material ones are resolved (or explicitly accepted with reason) before publish.
- This is a step inside the existing 准出, not a third blocking gate — it runs out-of-band like the governance behavior-audit, and its output is advisory input to the compliance Gate-2 decision, not a separate gate.
Rules
- A decision missing any of the 6 fields (§A) → the
solution_documentis incomplete → 准出 (Gate 2) fails. - The red-team pass (§B) must have run before publish; its findings + resolutions are recorded in the task ledger.
- Decoupling (skill_spec §9): generic. Which model(s) red-team and which decisions are "key" are
instance/project values — e.g.
project_profile.yamlmay name the red-team reviewer model; not hardcoded here.