AgentProvenance v0.8.1

English | 中文

An external-evaluator reference integration, demonstrated with Jev. This patch-sized addition lives in demo/jev-judge/, outside the core capture and query paths. It demonstrates the existing evidence/signal contract rather than adding a built-in AI analyst or workflow platform.

What is added

• A direct TypeSafe API example with three typed decisions: declared action, runtime conformance and evidence of unauthorized secret transmission.
• A separately launched demo workbench displaying the original evidence and answers, a question/criteria diff, comparison metrics and a human review journal.
• Manually authored baseline and candidate rubrics evaluated on identical case states. A fixed coverage guard is applied equally to both versions.
• Explicit candidate approval/rejection and reviewed quality_signal export using existing AgentProvenance interfaces. Approval becomes stale if a reference review changes; no policy is installed or automatically promoted.

The core owns evidence, correlation, verification and signal ingestion. This example owns its Jev calls, rubrics and review procedure. It adds no main CLI command, daemon worker, main dashboard navigation or mandatory provider setup.

Start with the Jev demo. It verifies the existing signed multi-agent bundle, makes twelve explicitly authorized API calls and saves a private local study. Reopening it requires neither a key nor more model calls.

Validation and limits

On 2026-09-22, all twelve live calls returned jev-1.13.0. The refusal case's raw transmission answer changed from not_observed in v1 to unknown in v2; other case classifications stayed the same. Both versions remained unknown after the deterministic missing-coverage guard. This selected comparison is not a held-out benchmark, a model-training result or proof of general safety.

The source bundle's signature and graph were checked. Local browser acceptance used a separate copy for review, approval, signal download/import/readback and graph verification. Test reviewer records are explicitly marked as automated; the primary study is left pending real human review. Keyless tests cover the client and review boundaries; the CI job does not send evidence to a provider.

Review names are self-reported. Local hashes detect accidental changes against the stored manifest, not a malicious owner's rewrite or rollback of all local files. The source signature does not sign subsequent judgments or approvals. This is a loopback demo, not an authenticated multi-user approval service.

Compatibility

No core Go package, database migration, original demo bundle or existing dashboard route is changed. The Jev UI is separate from the read-only evidence dashboard. Python 3.9+ on macOS/Linux is required only for this optional demo; Windows users can run it under WSL. Existing capture and replay workflows are unchanged.