Trace patterns

After the study migration is verified for your workspace, open retained studies from Investigations. Existing links keep their study, example and filter context. Home remains a read-only study view; Trace study review keeps its editing permissions. Check candidates remain available from the Investigations collection. A rollback restores the earlier navigation and keeps saved studies and reviews.

This workflow is available only to enabled workspaces in production. If its navigation item is absent, ask your workspace administrator to check access.

Watch the workflow: Review a proposed trace check · Authored dummy data

The video shows the earlier Trace study navigation. Migrated workspaces open retained studies from Investigations. Check candidates remain a separate workflow.

Inspect authored study coverage, eligible counts and a saved example. Unknown evidence and correlation remain explicit.

Read the walkthrough transcript

• 00:00 — This study uses invented examples. Source completeness is unknown.
• 00:05 — Read the eligible count, then compare the three evidence results.
• 00:11 — Read the saved sequence. This attempt ended without a saved completion.

The Trace patterns page groups possible gaps by the job an agent was doing. Choose a pattern type from the filter, then select a finding. Each row shows the finding, its use-case tag, the eligible and sampled turns, and the number of distinct turns where the recorded signal appeared. One turn can appear under several patterns, so the row counts should not be added together.

For example, a document-review agent might ask a picture tool to locate a signature area. An early attempt can fail, then a later view can work. The trace page should show both attempts. The failed attempt is a useful review signal; it does not, by itself, prove that the agent missed a signature or that the customer's work failed.

Finding names should say what work was underway. For example, a transaction compliance review checks paperwork against a checklist, while a transaction paperwork check compares uploaded documents with the stored record. If a turn has no successful saved completion, the page says exactly that. It does not claim the work was never completed later or that the customer's final result was wrong.

Select a pattern to open its right-side drawer. The list remains in view on wide screens, while the drawer shows the use case, explanation, signal, comparison, unknown, and eligible totals, plus a few saved turns. On a narrow screen, the drawer fills the screen with a clear close action. Choose a turn for a short recorded-sequence preview, then return to the pattern or open the full review.

Review all saved examples opens a full-width work view. Its table shows each turn's trace result, time, and one retained tool call when available: a failed call first, otherwise the first call. That column is a route into the evidence, not proof of why the task succeeded or failed. Filter by signal, sort by a column heading, or move between pages; select a turn to see why it was included and what the trace recorded. The questions used to inspect trajectories and the source-study details are in expandable sections. A source link appears only when the recorded identity resolves to that execution. If the trace cannot show the page, human decision, required hook, or final outcome, the page says what is still unknown.

An authorized reviewer can mark an example correct, incorrect, or cannot tell yet, with a reason. They can also suggest a better use-case tag or describe how a proposed classifier should change. These actions save review feedback; they do not silently change an approved classifier. A revised check still needs a versioned draft, comparison on eligible examples, independent outcome review where needed, and owner approval before official scoring.

Draft classifier runs save their exact model answer for reuse. When the customer, check version, model, questions, and evidence match a previously completed request, the runner can reuse that answer instead of calling the model again. Saved classifications support filtering and counts across the reviewed traces. A changed definition or evidence can require a new model call. Draft results remain separate from approved production results.

See Evaluator packs for the full review and approval lifecycle.