Self-Modification Governance Implementation Checklist

This checklist translates the governance architecture in self-modification-governance-byline.md into concrete engineering work items.

Scope and assumptions

• Runtime: SWARM orchestrator + governance engine.
• Goal: ship a minimally auditable two-gate modification loop before adding full compositional simulation and advanced rollout automation.
• Current calibration proxy mapping:
• tau_min proxy -> governance.refinery_p_threshold
• K_max proxy -> governance.memory_write_rate_limit_per_epoch

Phase 0: Hardening prerequisites

• Define trust boundaries in code ownership:
• immutable governance policy surfaces
• mutable agent/runtime surfaces
• Add signed policy bundle loading path (hash + signer + version).
• Add policy-hash and artifact-hash fields to run metadata.
• Add failure mode: any attestation mismatch blocks promotion.

Phase 1: Byline provenance foundation

• Define canonical Byline schema (JSON) with:
• identity fields (event_id, agent_id, run_id, timestamp)
• decision context references
• verification outputs
• lifecycle state transitions
• integrity chain fields (entry_hash, prev_hash, signatures)
• Implement append-only storage API.
• Implement strict lifecycle transition validator.
• Add replay tool: reconstruct full modification history by run_id.
• Add tests:
• schema validation
• transition invalidation
• hash-chain integrity detection

Phase 2: Gate 1 (tau_min) implementation

• Implement deterministic tau computation from held-out metrics.
• Add metric-direction registry (higher-is-better vs lower-is-better).
• Add confidence-adjusted margin term (eps_j treatment).
• Implement policy thresholds by risk tier.
• Log gate decision artifacts:
• raw metric values
• normalized margins
• final tau and threshold
• Add tests:
• deterministic replay consistency
• threshold boundary behavior
• fail-closed on missing metrics

Phase 3: Gate 2 (K_max) implementation

• Define deterministic complexity-weight model for modifications.
• Implement rolling-window budget accounting.
• Enforce cap with explicit outcomes:
• permit
• deny
• escalate-to-human
• Add per-agent budget telemetry and reset/consolidation policy.
• Add tests:
• cap exceed behavior
• rolling-window expiration behavior
• no negative/overflow accounting

Phase 4: Deterministic risk-tier classifier

• Implement rule-based classifier over:
• file/path surfaces
• permission policy diffs
• tool manifest diffs
• dependency diffs
• Emit signed classifier report with matched rules.
• Add anti-bypass behavior:
• unknown scope -> stricter tier
• classifier failure -> stricter tier
• transitive ownership escalation
• Add regression fixtures for tier assignment.

Phase 5: Promotion workflow and rollout safety

• Implement staged path:
• sandboxed test
• shadow
• canary-1
• canary-2
• full promotion
• Implement automatic rollback triggers with fixed thresholds.
• Implement kill-switch:
• stop promotions
• freeze affected populations
• open incident record
• Add incident runbook and rollback SLO checks.

Phase 6: Calibration and reproducibility

• Check in calibration script and runbook command.
• Store calibration artifacts under runs/<timestamp>_tau_k_calibration/.
• Require:
• runs.csv
• summary.json
• recommendation.json
• Pin seed list and scenario in docs.
• Add smoke check that verifies both gates were exercised:
• non-zero gate-hit count for K_max arm
• non-zero rejection delta for stricter tau candidates

Phase 7: Release criteria

• Byline completeness >= 99.9% for modification events.
• Deterministic replay success >= 95% on sampled events.
• Mean rollback latency < 10 minutes in fault-injection tests.
• No unresolved critical governance incident older than 24 hours.
• Documentation updated:
• architecture doc
• operator runbook
• calibration instructions

Current calibration snapshot

Latest run (seeded sweep):

• Artifacts:
• runs/20260214-020518_tau_k_calibration/runs.csv
• runs/20260214-020518_tau_k_calibration/summary.json
• runs/20260214-020518_tau_k_calibration/recommendation.json
• Recommended values from that run:
• tau_min = 0.55
• K_max = 6

Reproduce:

python scripts/calibrate_tau_k_memory.py