Attempt trace controls
| View | What it answers |
|---|---|
| Agent output | The files the Attempt wrote to its output folder, with an artifact inspector for previewing, comparing, and downloading them alongside screenshots and Verifier evidence. |
| Screenshots | Every image a tool returned, in order; each one jumps to the call that produced it. |
| Agent / Runtime / Verifier | Switches the timeline between the recorded agent conversation, sealed Runtime events, and sealed Verifier events. |
| Verdict | Score, each criterion with its reasoning, task prompt, submission, persisted Verifier records, and re-verification history. |
| Details / Raw JSONL | The selected record, harness, model, usage, tools, sessions, provenance, available Artifacts, and the raw stream; Raw JSONL downloads the complete sealed event stream. |
| QA | Review notes attached to the Attempt or selected event. |
The sealed evidence boundary
Evidence-mode Verifiers run over frozen evidence after the Attempt is finalized. They do not join the rollout network or inspect whatever the live Environment happens to contain later. That separation is what makes re-verification reproducible.
Attribution comes from the channel
A record belongs to the seat, verifier, or producer channel that delivered it. Fleet does not trust a payload merely because it claims another identity. Tool capability is also enforced by the server-side channel allowlist, not prompt wording.
External trace states
| State | Meaning |
|---|---|
| Ready | The external producer trace was retrieved and can be inspected. |
| Pending | A normally short-lived asynchronous indexing state. The trace pane shows the source reason and automatic retry timing; persistent Pending after the Attempt is sealed is a trace-delivery or indexing diagnostic, not evidence that execution continues. |
| Missing | The sealed Attempt declares no externally exported producer-trace session, so no external producer trace is available. |
| Error | The trace source failed; inspect the source-specific diagnostic. |