You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Training metadata includes the dataset, random seed, test split, quality metrics, model version, artifact path, SHA-256 digest, signature, public key, source revision, and builder identity. Local runs use `source_revision=local`; GitHub Actions records `GITHUB_SHA`.
@@ -194,7 +195,7 @@ flowchart TD
194
195
195
196
## Roadmap
196
197
197
-
Phase 2 has started with a read-only investigator contract integrated into audit records and covered by end-to-end release scenarios. Its output is structured, evidence-cited, schema-validated, and advisory, with missing and contradictory evidence flagged explicitly. An optional local Ollama transport is timeout-bound and receives structured evidence only. The intelligence layer must match the deterministic decision, has no unrestricted shell or Kubernetes access, and is not part of the final release authority.
198
+
Phase 2 has started with a read-only investigator contract integrated into audit records and covered by end-to-end release scenarios. `scripts/investigate_release.py` can regenerate an advisory investigation report from an existing audit log. Its output is structured, evidence-cited, schema-validated, and advisory, with missing and contradictory evidence flagged explicitly. An optional local Ollama transport is timeout-bound and receives structured evidence only. The intelligence layer must match the deterministic decision, has no unrestricted shell or Kubernetes access, and is not part of the final release authority.
0 commit comments