A complete walkthrough of the Adaptive Governance Framework v3.2 — the 5-layer system powering copilot-kit.
The Adaptive Governance Framework (AGF) v3.2 is a risk-activated enforcement system embedded in .github/copilot-instructions.md. It runs silently on every GitHub Copilot request, scoring it for risk and adapting the response format, strictness, and validation accordingly.
The AGF has 5.5 layers (Layer 4.5 is the Response Format Contract). Each layer builds on the previous, creating a coherent system from safety rules to output schemas.
Meta Principles
Always Active · Non-Negotiable
10 immutable anchors that no other rule can override. Covers safety (never produce harmful output), scope (do exactly what is asked), security (never hardcode secrets), quality, reproducibility, and language mirroring.
Risk Engine
Silent · Per-Request
Every request is scored on 4 axes (0–3 each), producing a total score 0–12. The score maps to a risk level: L0 Trivial, L1 Routine, L2 Elevated, or L3 Critical. This computation is internal — Copilot never narrates it.
Strictness Modes
Auto-Set by Risk · User-Overridable
Four operational modes that control response format, Socratic questioning, agent banners, and validation checklist scope. Auto-set by the Risk Engine but overridable with plain English commands.
Conflict Resolution
Priority Tiers P0–P4
When rules conflict, the Conflict Resolution Matrix determines the winner. Priority tiers are: P0 (governance rules) > P1 (agent frontmatter) > P2 (skills) > P3 (workflows) > P4 (user request). Safety rules at P0 are absolute.
Response Format Contract
Output Schema by Risk Level
The output schema adapts to the final risk level. L0–L1 uses free-form prose. L2 requires 6 structured fields. L3 requires 10 fields including an explicit rollback plan and Final Safety Confirmation.
Validation Feedback Loop
Post-Action
After generating a response, Copilot scores it against the active Meta Principles and Strictness Mode. Violations surface as inline notes (L0–L1), Quality Flags (L2), or full blocks (L3). Breaking changes auto-elevate the score by +3.
The scoring is silent
Learn about risk levels