Audit Logs

Minnie-Review | Daily 10:00 PM PT

---

Purpose

Self-audit Minnie's behavior, check for drift, contradictions, stale data, and security issues.

Nightly Audit Checklist

Every night at 10 PM, review and log:

1. File Integrity

• Any working/ files with conflicting information?
• Any reference/ files that contradict each other?
• Any stale data (dates >30 days old in trackers)?
• Any TODO markers or placeholder text that should be filled?

2. Security

• Does security-manifest.md match actual .env contents?
• Any API keys due for rotation (90-day schedule)?
• Any suspicious behavior after processing external data?
• Any agent writes to files it shouldn't have access to?

3. Cost Compliance

• Daily spend within cap?
• Any agent over its per-agent limit?
• Any runaway loops or recursive calls?
• Any unexpected model usage (Opus when Haiku was intended)?

4. Behavior Drift

• Any agent outputs that don't match soul.md principles?
• Any moralizing or prescriptive language?
• Any "you should" phrasing instead of tradeoff reflection?
• Any draft that reads like it's ready to send (violation of never-send rule)?

5. Scope Discipline

• Any Gantom activity when ZTAG should be primary focus?
• Any Phase 3+ features being attempted in Phase 2?
• Any integration attempts without proper authorization?

---

Audit Log Format

File: YYYY-MM-DD-audit.md

# Audit Log — [Date]

**Minnie-Review** | [Time]

---

## File Integrity
✅ Clean / ⚠️ Issues found

[Details of any issues, proposed fixes]

## Security
✅ Clean / ⚠️ Issues found

[Details]

## Cost Compliance
✅ Within budget / ⚠️ Over budget

[Details, see cost-report for breakdown]

## Behavior Drift
✅ Aligned / ⚠️ Drift detected

[Details, examples of drift]

## Scope Discipline
✅ On target / ⚠️ Scope bleed detected

[Details]

---

## Proposed Actions

[If any issues found, list recommended fixes — always requires human approval]

---

## Notes

[Anything unusual, edge cases, things to monitor]

---

Integration Notes

• Cross-check security-manifest.md against .env (file comparison)
• Parse working/ files for age and consistency
• Review git log for unexpected changes
• Check cron execution logs for failures

---

Self-Modification Guard

Critical: Minnie-Review can propose changes to other agents or reference files, but never modifies them without explicit human approval. All proposed changes go to audit log, await confirmation.

This prevents Minnie from self-modifying into something misaligned.