Task verdicts and evidence assessments

Watch the workflow: Follow an authored proxy signal to its recorded trace · Authored dummy data

Filter one authored Task Success proxy in Tensile, read its recorded explanation and inspect its session trace. No evaluator or provider was run for this fixture. The exact historical input and rubric are unavailable, and the signal is not a judge result or customer outcome. No feedback is saved.

Read the walkthrough transcript

• 00:00 — Filter an authored Task Success proxy in Tensile. This signal is not a judge result.
• 00:05 — Open the recorded proxy. Its exact historical input and rubric were not recorded.
• 00:11 — Read the authored explanation. No evaluator or provider was run for this example.
• 00:18 — Follow the recorded session trace. Two events do not establish customer success.

In Evals, open a Task Success row or its details button. The inspector shows the recorded judge classification or neutral proxy signal, its score and source. The recorded result is preserved. A judge pass does not confirm customer success.

Read the evidence assessmentCopy Anchor Link

When a stored assessment is available, read its requested outcome, observations, evidence, open questions and next check. The assessment names its actual model and remains separate from the recorded verdict. It can agree, disagree or find insufficient evidence; disagreement does not change the recorded classification. Opening or refreshing the inspector does not call a model.

Use Sources under a claim to read its quoted input or output. Each passage has its own source and passage label, including multiple quotes from one source. A statement by the agent and a tool acknowledgment have different evidential meaning from a confirmed destination result. Copy includes only the currently authorized assessment and references. View session opens the recorded source session when its timestamp is available.

If an assessment is missing, expired or inaccessible, the inspector explains that state without inventing an answer. The recorded result and verdict feedback remain available when that exact verdict is still accessible. Reopening or refreshing checks current permissions and source availability; a source outage can temporarily withhold an assessment.

Save feedbackCopy Anchor Link

Verdict feedback appears at the top, above the evidence assessment. Use the check to agree with the recorded classification or the cross to disagree. The information control opens optional feedback details. Accessible labels and hover/focus hints identify each action. Use Was this explanation useful? after reading the assessment to give thumbs up or down. Each signal saves immediately and has its own saved state. Optional feedback adds a reason, your review commentary, and—for verdict feedback—a proposed outcome. Reference existing cited records rather than copying source text into a note.

Feedback edits append review history. Proposed outcomes remain unadjudicated: they do not change the verdict, create training labels or become verified ground truth. Stored notes are visible to their author and workspace owners/admins; explanation notes also require continued access to that exact assessment.

Inspect versions and calculationsCopy Anchor Link

Shared links retain the exact verdict version. A retained historical version is read-only and does not select a newer row as the same result. If that version is no longer retained, the inspector says it is unavailable. A separate action can open the currently listed version of the same verdict.

Expand How this score was calculated for full recorded numeric precision and the producer’s saved rationale. Rounded values cannot establish equality with a cutoff; a cutoff missing from the historical record remains unknown. Configuration history and today’s definitions remain separately accessible under the source and review details.

On small screens, the inspector replaces the list; use Back to list to return. Keyboard users can open the details button with Enter and close the inspector with Escape.