| name | self-evolution |
| description | Governance lifecycle manager for the Three-Layer Governance OS. Six operations: CLASSIFY, CAPTURE, PROMOTE, REFINE, RETIRE, COMPRESS. Manages principles (SOUL), rules (AGENT/STEERING), and gates (hooks/code). Detects capability gaps, captures corrections with bias tags, auto-promotes recurring patterns (3x threshold), and prevents governance bloat.
TRIGGER: "add rule", "steeringify".
NOT FOR: for one-off file edits (just edit directly). |
| tier | always |
| platform | all |
Self-Evolution — Governance Lifecycle Manager
Manages the Three-Layer Governance OS: Principles → Rules → Gates.
Read INSTRUCTIONS.md for the full protocol.
Six Operations
| Operation | Trigger | What It Does |
|---|
| CLASSIFY | Any governance change proposed | Classify layer, parent principle, check conflict/budget |
| CAPTURE | User correction detected | Write C-entry with bias tag (A/B/C/D) |
| PROMOTE | Pattern seen 3x OR user approves | Elevate from EVOLUTION → AGENT/STEERING (absorbs steeringify) |
| REFINE | Principle not precise enough | Sharpen existing principle/rule wording |
| RETIRE | Rule covered by gate, or 30d idle | Remove from active rules, record in EVOLUTION |
| COMPRESS | 3+ rules share same root | Merge into single rule with broader coverage |
Trigger Detection (unchanged from v1)
🔴 Reactive — Capability gap
🟡 Proactive — Better way exists (deferred)
🟢 Correction Capture — User pushback → C-entry with bias tag
🔵 Stuck — Going in circles
Priority: Stuck > Reactive > Correction > Proactive
Limits: Max 3 triggers/session. 60s cooldown.
Bias Classification (on every correction)
| Bias | Pattern | Indicator |
|---|
| A | Premature completion | "Good enough", skip review, defer known issue |
| B | Inference over evidence | Assert without reading, stale memory |
| C | Productivity over quality | Rush output, shallow analysis |
| D | Surrender on obstacle | Give up, ask user to compensate |
Governance Budget (enforced on PROMOTE/REFINE)
- SOUL.md principles: ≤12
- AGENT.md rules: ≤25
- STEERING.md standing rules: ≤15
At cap → COMPRESS or RETIRE existing before adding.
Intake Classification (on every governance change)
Before writing to SOUL/AGENT/STEERING:
- Layer: principle / rule / gate?
- Parent: P1-P7?
- Conflict: duplicates/contradicts existing?
- Budget: current count vs cap
- Source: user / correction / pipeline / self-detect
Output brief → user confirms (user source) or auto-apply (3x agent evidence).
Config
| Key | Default | Effect |
|---|
enabled | true | Kill switch |
max_triggers_per_session | 3 | Per-session cap |
promotion_threshold | 3 | Same-class corrections needed for auto-promote |
deprecation_days | 30 | Days idle before retirement candidate |
Writing to EVOLUTION.md
Use Read + Edit tools. Never python3 locked_write.py (PyInstaller crash).
JSONL changelog append after every mutation (mandatory).
Verification
