A score does not survive examination. A signed packet does.
Most eval tooling stops at a number. This walkthrough follows one release from the eval suite to the signed gate that held it — every reviewer, override, and approval on one chain you can hand to an examiner. When the EU AI Act provenance clock enforces in August 2026, this is the record they ask for.
The eval suite, the target model, the policy gate, and the release path enter one record.
Every edge carries who reviewed it, who held it, and why the run moved forward.
The signed release packet an examiner can read — owners, approvals, and exports on one chain.
Status, owner, and evidence travel with every edge.
142 cases
0.94 pass
12 routed
1 warning
hold
signed
What the graph proves.
A regression escapes. Trace it back through every reviewer action and override to the run that shipped it.
A gate holds the release. The hold, the owner, and the escalation are recorded — not buried in a chat thread.
An examiner asks who approved the model. You hand over a signed packet, not a reconstruction.
Read-only here. Yours in a pilot.
The figures here are illustrative seed data. In a pilot, the graph reads from your own evals, your own reviewers, and your own release holds — and the signed packet at the end is one you can defend.