Organizations building AI products discover that ungoverned agent systems do not survive contact with production, audit or regulation. Our answer is an architecture in which six responsibilities remain separate, auditable and combinable. Never merged.
The six responsibilities
- TruthMetal defines and versions expected truth and evidence: oracles, golden cases, evaluation specifications, provenance and revocation.
- Verentir measures and judges independently: it runs evaluations, attributes failures and detects regression. It recommends; it never enforces its own verdict.
- Thalamus applies AI runtime control: policies, model routing, limits, guardrails and interruption. Every governed call passes through it.
- Agent Orchestration coordinates execution: missions, roles, sequencing, retries and explicit stop criteria.
- RBX Governance preserves institutional memory: decisions recorded with rationale, ownership and impact. Git writes and approves; systems and interfaces are read projections.
- Strategos presents the situation and supports human judgment: projections, dossiers and strategic triggers always authorized by people.
The invariants
- Recommendation and enforcement are distinct powers. The plane that measures never enforces; the plane that controls never grades itself.
- Consequential action is human-gated: spend, commitments, external claims and strategic triggers.
- Evidence over assertion. A trace or a dashboard, alone, is not an audit record.
- An indeterminate outcome is never blindly retried; it is reconciled first.
The canonical version of this manifesto is versioned in our governance repository, with change control by recorded decision. Nothing here changes silently.