eval-agentic-cleanup

To audit agentic eval runs, perform a read-only completeness and health check that verifies job completion, score presence, and trace linkage by parsing per-trial result.json files without mutating data.

Remediation of missing HF traces involves an idempotent flow that performs manual trace uploads, Supabase DB registration, and trace dataset uploads, followed by cross-user foreign key safety checks.

To fix invalid scores and recover failed trials in SLURM-based workflows, apply targeted re-run capabilities to resume only errored trials and require re-running the audit to confirm success.

Yes, the agentic eval remediation flow implements cross-user foreign key safety checks during manual DB updates and trace re-registration to prevent data integrity issues across different users.

Yes, you can use the targeted re-run capability to resume only errored trials instead of the entire evaluation job, then re-run the read-only audit to verify completeness and trace registration.

An evaluation audit may show missing traces after a completed SLURM job if Hugging Face traces or Supabase DB entries failed to register properly, requiring manual trace upload and DB registration remediation.