audit-distill

Audit-Distill — From Audit Results to Framework Knowledge

Register: TECHNIQUE Goal: Transform VHEATM audit findings into structured Super Skills knowledge that shapes future cycles. Constraints: Only T1/T2 evidence enters Gotchas sections — no speculation. Tag every M.AT entry with QBR formula version. Deduplicate before writing. Adapt: extraction depth to audit size; a 5-finding audit needs 10 minutes, a 30-finding audit needs an hour.

The feedback loop this closes:

VHEATM audit → audit-distill → Gotchas + PATTERN-DEBT + M.AT seed
     ↑                                        ↓
Better audits ← Better agent coding ← Better framework knowledge

No other agentic coding framework has this loop. Guard it.

Announce: "Using audit-distill to extract [project-name] audit findings into Super Skills knowledge."

---

Pre-Check

# Confirm KB structure exists:
!ls docs/superskills/ 2>/dev/null || echo "⚠️  Run skill-init first"

# Check existing Gotchas content (don't duplicate):
!grep -r "^\- \[20" ~/.claude/skills/*/SKILL.md 2>/dev/null | wc -l | xargs echo "Existing Gotcha entries:"

# Idempotency: check if this audit was already distilled
!grep -r "[audit-source-name]" docs/superskills/compound-wiki.md 2>/dev/null && echo "⚠️  Already distilled — check for duplicates"

---

Step 1: Parse and Classify Findings

Read the audit report. For each finding, extract:

FINDING:
  id:          [H-01, SEC-001, MANDATORY-01, etc.]
  type:        [BUG | SECURITY | PERFORMANCE | OPS | COMPLIANCE | AI-SAFETY | ARCHITECTURE]
  qbr:         [score] (note formula version: v8.0-additive | v16.x-multiplicative)
  severity:    [MANDATORY | REQUIRED | RECOMMENDED | LOW]
  actual_outcome: [if known post-fix: INCIDENT | NEAR_MISS | NO_IMPACT | OVERESTIMATED | UNKNOWN]
  evidence_tier: [T1 | T2 | T3 | T4]
  description: [1-2 sentence summary]
  root_cause:  [structural cause, not instance]
  lesson:      [what to do differently next time]