Coherence System Atlas
A read-only system atlas of every engine, phase, capability, decision, and evidence record in the Coherence Engines control layer. Production wiring is OFF by policy; this is internal architecture documentation only. Access is server-validated against an internal-role allowlist.
Access Control
Authoritative boundary: server-validated authorizeInternalAtlasAccess backend function checks user.role against an internal-role allowlist that is NOT shipped to the client.
Trusted internal roles: internal_owner, internal_superadmin, coherence_internal_architect.
Default: deny. Unknown, missing, stale or malformed role → deny. Authorization service unavailable → deny. Generic admin role → deny.
The manifest is lazy-loaded only after internal_access_allowed === true; it never loads for unauthorized users.
Executive Overview — Non-Technical
Coherence Engines is an independent control layer around AI systems that determines:
- which model may be used, under which conditions;
- which actions an AI agent may take;
- whether the result is reliable;
- whether a failure can be repaired;
- when a person must review;
- when the system must stop;
- how evidence is recorded.
It is not the AI model itself — it wraps around AI systems to provide control.
Executive Overview — Technical
The control layer is composed of orchestrators (Phase 72, Phase 79), registries (proposal, permission, capability, asset), gates (role separation, capability delta, semantic buffer), evaluators (model evaluation integrity, goal drift, candidate), ledgers (improvement, provenance), routers (vision, operational fabric), verifiers (Phase 81 tool result), repair engines (bounded tool repair), drift monitors (Phase 85 behavior drift watch), sandbox assurors, improvement lifecycle (Phase 86), authority precedence (Phase 72 final, Phase 80 tool, Phase 81 verification, Phase 85 advisory, Phase 86 advisory), and clear integration boundaries (no live calls, no customer data, production wiring OFF).