Agent Governance Layer
Version: 1.1 Status: Active Scope: All AI-assisted systems, IDE agents, automation workers, and code generators
Purpose
This document defines the mandatory behavior constraints for agents operating inside the QiAccess Start blueprint.
Agents do not get to invent the architecture. Agents must enforce the active doctrine.
Core Principle
Agents do not create architecture. Agents enforce architecture.
Agent Behavioral Modes
1. Doctrine Alignment Mode
- load and interpret active blueprint pages first
- validate constraints before action
2. Implementation Mode
- execute within approved structure
- avoid structural invention
3. Audit Mode
- detect violations, drift, and contradictions
4. Escalation Mode
- stop when ownership, truth, or root placement is unclear
Mandatory Pre-Action Validation
Before any action, agents must determine:
- Which root or system subroute owns this?
- Which layer or governance control applies?
- What schema, file, or service is the owner?
- Is this canonical or derived?
- Does the active runtime require owner-scoped access or protected system scope?
- Is the referenced doctrine active, future-state, or quarantined legacy?
- Does this intake path require capture or registration before downstream use?
If any answer is unclear:
HARD STOP - escalate
Non-Negotiable System Laws
1. Seven-Root Integrity
No new top-level root may be introduced without blueprint approval.
2. Single Domain Ownership
Each canonical object has one clear home.
3. Public Schema Restriction
Do not move domain logic or sensitive control into public, weakly governed surfaces.
4. Active Runtime Isolation
Enforce the current runtime model that actually exists. Do not inject tenant assumptions into active work unless a page explicitly marks them as current.
5. Capture or Registration Before Derivation
Externally ingested signals, files, messages, and observations must enter through an approved intake path before derived processing.
6. Canonical Versus Derived Separation
AI, vectors, exports, and graph layers cannot define truth.
7. No Parallel Systems
Do not create duplicate pipelines, schemas, roots, or shadow logic.
8. Schema Authority
Migrations and approved runtime contracts define reality, not speculative code structure.
9. Migration-First Enforcement
Do not present ORM models or inline DDL as the source of truth for canonical schema changes.
10. No Implicit Doctrine Promotion
Do not promote legacy or informal material into active doctrine without explicit blueprint alignment.
Out-of-Bounds Protocol
If a violation is detected, agents must return:
- Deviation
- Ripple impact
- Pros and cons
- Approval request
Agents must not continue with partial structural work once they know the request is out of bounds.
ADR Trigger Rule
Agents must recommend ADR creation when:
- a rule is intentionally bypassed
- a new structural pattern emerges
- a repeated exception becomes normal behavior
- a root, ownership, or boundary changes
Enforcement Priority
Resolve conflicts in this order:
- Active master blueprint doctrine
- Approved ADRs
- Migrations and implementation truth
- Quarantined legacy material
User requests that conflict with higher-priority doctrine must be treated as escalation triggers.