Most teams say their AI systems are “auditable.” Very few can hand you a folder that proves it.
This packet is that folder — a worked example of the evidence bundle a governed agent system should be able to produce for one single correlated decision. Everything inside hangs off one correlation ID, the way it would in production.
What’s inside
corr-<id>/
├── decision_traces.jsonl # what the agent did, step by step
├── evidence.jsonl # the evidence each step relied on
├── policy/
│ ├── decision.rego # the policy the decision was evaluated against
│ └── metadata.json # policy version and provenance
├── schemas/
│ ├── decision_trace.avsc # versioned contract for the trace itself
│ ├── evidence.avsc # versioned contract for evidence records
│ └── SHA256SUMS # integrity hashes
├── observability/
│ └── pointers.json # links into logs/metrics for this run
└── supplychain/
├── cosign.pub # signing key for verification
└── sbom_out/ # SBOMs for the services involved
Who it’s for
Platform and engineering leaders taking agentic systems into environments where “trust me” isn’t an acceptable answer — regulated industries, high-consequence operations, anywhere a board, auditor, or regulator can ask show me.
How to use it
- Read the trace first.
decision_traces.jsonlis the spine — each line is one step of delegated judgment, bound to policy and evidence. - Diff it against your own system. For each file, ask: could we produce this today, for any decision, on demand? The gaps are your governance backlog.
- Steal the schemas. The Avro contracts are deliberately minimal — versioned, hash-summed, and boring. That’s the point.
The packet is ungated and free to reuse internally. If you want the thinking behind it, the Trust pillar holds the argument and the DecisionOps Control Plane shows the architecture it comes from.