2.6 KiB
2.6 KiB
Agent Learning Policy
Status: planned / narrow v1 Scope: guardrails for recursive self-improvement
Prime Rule
PodMan may improve its agent behavior only when the improvement is narrow, evidence-backed, versioned, and reversible.
Allowed Learning
PodMan may learn:
- Which prompt version produces clearer interventions.
- Which detector threshold reduces false positives.
- Which routing channel gets accepted without being intrusive.
- Which verifier best predicts user acceptance.
- Which graph-discovery rule produces cleaner risk paths.
Disallowed Learning
PodMan must not:
- Promote a strategy because the model says it is better.
- Rewrite broad system behavior from one example.
- Hide failures, dismissals, or rejected candidates.
- Learn from raw screenshots, secrets, or private terminal content.
- Turn voice into the default route.
- Create irreversible actions without human approval.
Promotion Rules
A candidate strategy can become active only when all are true:
- It has a parent strategy version.
- It describes one concrete behavior change.
- It has a verifier plan.
- It has evidence from a run, outcome, or test.
- It improves or fixes the target metric.
- It does not increase user interruption without payoff.
Rejection Rules
Reject and retain the candidate when:
- The verifier regresses.
- The change is too broad.
- The evidence is missing.
- The candidate conflicts with privacy rules.
- The candidate makes the demo less stable.
Evidence Strength
| Evidence | Strength | Use |
|---|---|---|
| Model opinion | Weak | Proposal only |
| Trace observation | Medium | Candidate rationale |
| Human accepted outcome | Strong | Promotion candidate |
| Human dismissed outcome | Strong | Suppression or rejection |
| Automated verifier | Strong | Promotion or rejection |
| Repeated accepted exact signature | Strong | Policy confidence increase |
Versioning Rules
- Strategy versions are immutable after promotion or rejection.
- There is one active version per
podId + kind. - A rollback activates the previous version; it does not edit history.
- Parent-child lineage must be preserved.
Safety Rules
- Store summaries, not raw sensitive content.
- Prefer deterministic checks over model judgment.
- Use exact MongoDB recall before vector recall.
- Ask for approval before changing code or data with external effects.
- Treat hackathon demo stability as a hard constraint.
Demo Honesty
Seeded strategy versions are acceptable when labeled as demo-backed. Do not claim a strategy was learned live unless a run and outcome actually created the promotion evidence.