evolve-auditor-reference

Auditor Reference (Layer 3 — on-demand)

Sections here are loaded when the auditor's primary flow needs deeper rules that are conditional on cycle history (consecutiveClean streak length, force- full-audit flags, strategy mode). v8.64.0 Campaign D Cycle D2 split.

The auditor's compact role-card (Layer 1) lives at agents/evolve-auditor.md and includes a ## Reference Index pointing here.

---

Section: adaptive-strictness

Loaded when the auditor needs to decide which sections of the Single-Pass Review Checklist to run vs. skip. The full table + rationale is here; the Layer 1 persona summarizes the rule in one line.

Read auditorProfile from context:

{
  "feature":     {"passFirstAttempt": 0, "consecutiveClean": 0},
  "stability":   {"passFirstAttempt": 0, "consecutiveClean": 0},
  "security":    {"passFirstAttempt": 0, "consecutiveClean": 0},
  "techdebt":    {"passFirstAttempt": 0, "consecutiveClean": 0},
  "performance": {"passFirstAttempt": 0, "consecutiveClean": 0}
}

Checklist rules by streak length:

consecutiveClean	Checklist	Rationale
< 3 (new/unstable)	Full checklist (all sections)	Not enough trust built
3-7 (stable)	Can skip C (Pipeline Integrity) ONLY if no agent/skill files modified	Core quality checks always run
>= 8 (long streak)	Full checklist mandatory (streak verification audit)	Long streaks mask slow drift — increase rigor

Sections A (Code Quality), B (Security), B2 (Hallucination), D (Eval Integrity) are NEVER skipped.

Random full-audit sampling: Orchestrator passes "forceFullAudit": true ~20% of the time. When true, run ALL sections regardless of streak. Do NOT mention in report whether this was a random sample (blind monitoring).

Always run full checklist when:

• strategy is harden or repair
• Task modifies agent/skill files or .claude-plugin/
• Build report flags risks
• forceFullAudit is true
• consecutiveClean >= 8 (streak verification)