Gold Standard

Answer: "For a project that does THIS — are rules/standards/features 100% vs world-class? If not, what's missing?"

Four acts: AUDIT → FILL → ADOPT → CONFORM. Stop at any.

Manual /gold-standard = interactive setup: read references/wizard.md (dual-audience — layman 1-question default · programmer order→bill→pay). The auto/keyword path (the Triggers table) skips it.

Triggers

Consent gates (G1–G6, declared)

Hook cells assume an interactive session; non-interactive → D6, report-only: no gate fires. Manual /gold-standard wizard (references/wizard.md) carries its own setup gates — on-demand, not counted above.

Acts

ADOPT and every CONFORM fix: gated (G2/G3), never assumed (P2).

1. AUDIT — pick 3–5 named exemplars fresh at run time (the bar moves with the era — never reuse a remembered bar (P3)), derive the 100% checklist per dimension, score (✅/🟡/❌/N-A), give overall %. Previously filled/adopted rules are audit subjects too: a rule past its revalidate due date or contradicted by today's exemplars is a gap.
2. FILL — write missing MUST-HAVE rules into the project's rules home (.claude/rules/ → AGENTS.md → STANDARDS.md). Match project style + voice. Cite the exemplar. Invoke source-grounding for version-sensitive claims. Extend existing; never duplicate (P4). Check the retired-rules record first — never resurrect a rule retired with a reason unless the user overrides (P5). No overkill rules — only essential, practical, saturated ones (P6). Stamp each rule: <!-- coalmine: verified <YYYY-MM-DD> · exemplar <name> · revalidate <30|90>d --> — 30d fast-moving surfaces, 90d general; a CVE/advisory rule's event overrides its calendar stamp. Mechanics + why: references/method.md.
3. ADOPT — treat the completed ruleset as binding for the rest of the session. Code changes still need user approval — adoption governs how to work, not license to auto-edit (P7).
4. CONFORM — scan existing code against adopted rules; report violations (path:line · rule · evidence). Fix on approval: checkpoint → record a build+test BASELINE → one fix → build+tests → revert only if a NEW failure appeared versus the baseline. Style Drift: conform minority patterns to the dominant style (highest average usage); never start a standalone style refactor (P8).
5. RE-VALIDATE (runs inside every repeat AUDIT, or when offered on a past-due stamp) — verdict each CoalMine-stamped rule, all changes choice-gated (G4/P9):
   • still valid → re-stamp the date of what you re-read, touch nothing else (no churn; a file-level stamp over sections you did not re-read keeps its date, references/method.md Stamp date semantics);
   • stale but needed → rewrite against today's exemplar;
   • obsolete (subject removed, platform died, or substance merged into another rule) → delete the rule + record a one-line tombstone in the project's memory/decision log (retired <rule> <date>: <reason>) — dead rules burn context every session; the tombstone blocks the next FILL from resurrecting them.
   • CONSISTENCY (the agent trusts memory/rules it never verifies — so verify them): scan the memory/decision log and any in-repo rule register for (a) a prescribed fix/"decision" that contradicts a binding rule or another decision (e.g. prescribing randomness where a determinism rule forbids it) — a poisoned/stale entry; (b) references to a file, flag, or command that no longer exists. Flag each with the conflicting source quoted; correct only through the choice-gate (P9). This is the semantic half; the mechanical half (node scripts/consistency.mjs: cross-document counts, byte-identical doctrine mirrors, well-formed stamps) runs without an agent.

Exemplar-picking rules, scorecard mechanics, stamp/tombstone formats: read references/method.md before the first AUDIT.

Method

1. Bar — name 3–5 world-class exemplars + why them (cite real programs, not "best practices" (P11)).
2. Checklist — MUST-HAVE (table-stakes) vs EXCELLENCE (top-tier polish). Each tied to an exemplar.
3. Score — every criterion. 🟡 = half credit. N-A must be justified; unjustified N-A = ❌ (D2).
4. Gaps — prioritized: MUST-HAVEs first. Each: criterion · exemplar · move (the one rule or change that closes the gap) · effort · impact.

Dimensions (pick relevant)

Correctness · Security · Performance · UX/DX · Docs/onboarding · Testing/CI · Distribution/integrity · Observability · Governance/licensing · Maintainability · Compatibility · Error handling

Discipline

• State dimensions not assessed + why.
• Score/verify per P10–P12 + D1 (Prohibitions/Degrade paths below) — not restated here.

Prohibitions (P1–P16, declared)

Dedup: P5 restated at the RE-VALIDATE tombstone · P9 restated at the CONSISTENCY sub-bullet · P11 restated at Method §1 · P16 restated at the CLASSIFY-BLOCK write row — one row each, further mentions.

Degrade paths (D1–D8, declared)

D3 restated at four sites — the general clause, the Standard row's "(else single-agent)", the Heavy row's "if supported", and the Heavy-specific "escalate by model + reasoning only" — one row, four mentions. D4–D6 come from the shared Escalation footer below (Agent lane has no equivalent for D4/D5 — tier is asked, see G1, never degraded; D6 applies in either lane, wherever ask_question would fire). The Freshness cap (scope already audited this session → cap at Light) is a tier-selection modifier on G1, not a degrade branch — no capability lever is missing and there is no unhappy path, so it stays out of D. D7/D8 are this skill's CLASSIFY-BLOCK branches (skill-authoring.md §5b) — numbered here, not restated in the table below.

Grants & denials (CLASSIFY-BLOCK)

SPAWN/TIER/QUESTION-TOOL/NETWORK already discharged above — D3/D4/D6, D1. This table adds READ + WRITE; their denial branches are numbered into the Degrade paths ledger (D8, D7) rather than restated here.

Output — 5 locations, declared

(the AUDIT report only — FILL writes rules, RE-VALIDATE writes a tombstone, self error-report files an issue: none of those are part of this list)

1. Bar — category + named exemplars
2. Scorecard — | dimension | criterion | tier (must/excellence) | exemplar | status | evidence |
3. Per-dimension % + overall % (list N-A exclusions)
4. Gaps — criterion · exemplar · move · effort · impact
5. Verdict — 1 line + top 3 moves