Custody-First AI Governance: A Reference Architecture for Canadian Public Bodies

This paper specifies an open-source reference architecture for governing AI systems in Canadian public bodies, and fixes the protocol for validating it at Technology Readiness Level 4. The architecture answers a gap that current governance tooling leaves open: no product carries a clock that runs for as long as the consequences of a decision last. On their publicly documented capabilities, existing tools record that an assessment happened. They do not keep the assessment reviewable after the model, the reference data and the supplier have changed. The design is custody-first. Every registered AI system carries a decision class, a consequence horizon, an evidence horizon computed over named dependencies, and an accountable office. The Canadian federal rulebook (Directive on Automated Decision-Making, Algorithmic Impact Assessment, Policy on Service and Digital) is built in as versioned controls, not loaded as a plug-in, together with the AI Strategy for the Federal Public Service and the guiding principles for responsible AI. Every finding is a record in its own right: requirement, evidence slice, result, confidence, authority, timestamp. It is retained for the consequence horizon of the system it assesses. Reviewers can validate, reject, modify or annotate any finding, and an append-only log preserves the sequence of every decision. Erasure leaves a tombstone, never a silent gap. The paper doubles as the build specification. It fixes the data model, the control model, the traceability chain, the API surface and the sovereignty constraints. The reference implementation deploys on-premises with a single command and makes no outbound network calls. Section 10 fixes the lab validation protocol against a synthetic corpus of registered systems, written to the evidence standard that Innovative Solutions Canada requires for a Phase 2 bid. Section 11 sets out the path to TRL 5 through a pilot in a Canadian public body. Sovereign Digital Resilience series, paper 10. Draft v0.2, deposited under embargo pending v0.3 with independently witnessed validation results.