agent-trace-gov — risk-as-code demonstrator

Each row is one (workflow × risk-event), reconstructed from real agent traces. Click lineage for the evidence trail.

The first-paint hero: structural lineage COUNTED from traces (not a probabilistic cascade). RE-03 unauthorised external flows are lit; RE-08 multi-agent loops are dashed.

Authoritative taxonomy + framework citations from action_schema.json. Firing classes are highlighted; incident anchors live in the ATTEST precedent register, never on a finding.

The common-cause layer of the AGENT substrate — orchestrator, model provider, tool registry, classifier plane, credential broker. Blast radii are DERIVED from the same traces and lineage this page renders; each node names its rac analog (same mechanism, priced in dollars on the infrastructure substrate). Structural units only — atg deliberately carries no loss-dollar model.

Which regulatory obligations the firing risk-events trigger, grouped by jurisdiction. Article-level facts only — the materiality DETERMINATION is reserved to Legal/Compliance.

Every action requires a named human + ≥12-word justification (the A8 gate, enforced in your browser) and is hash-chained into the attestation log.

SHA-256 over canonical JSON, verified live in-browser — byte-matching the Python signer. Embedded attestations are illustrative demonstrator decisions; the chain math over them is real. Actions you take this session append locally and re-verify.

Real enforcement precedent keyed to the RE class (port spec §14.A.3) — deliberately separated from any single finding's magnitude.

The audit gates prove the classifier catches the shipped scenarios. This eval injects what it was never shown: in-class variants (different agents, tools, endpoints, boundary values) that must fire, and near-misses one predicate short of the rule that must not. Every figure hydrates from the sealed eval record.