Scores
Evaluate › Scores lists every judgment recorded in a tenant: who judged, the verdict, the answer it was about, and the note that says why. Judgments are recorded in the review queue or beside an answer in its conversation; this page is where they are read.
Who can open it
Section titled “Who can open it”| Role | Reads scores | Opens the answer a score judged |
|---|---|---|
| Admin | Yes | Yes |
| Reviewer | Yes | Yes, in its conversation, without traces |
| Auditor | Yes | Yes |
| Platform admin | Yes, in every tenant | Yes |
| Builder, End user | No | No |
Recording a judgment takes the Admin or Reviewer role; see Review queue.
What you see
Section titled “What you see”- The page header, with the tenant picker.
- The Judgments table, newest first, with Search judgments and a Judged by filter: Everyone, Person or Model.
| Column | What it shows |
|---|---|
| When | When the judgment was recorded. |
| Judged by | Person or Model. |
| Verdict | Acceptable, Wrong, Unsafe or Off policy. |
| The answer judged | The start of the answer. For an assistant answer, it is a link that opens the conversation at that answer. |
| Reviewer | Who recorded it: their email, or their account id when the sign-in carries no email. |
Expand a row to read Why, the note recorded with the judgment, or No reason was recorded.
When nothing has been judged, the page says Nothing judged yet and offers Open the review queue.
What you can do
Section titled “What you can do”Read why an answer was judged
Section titled “Read why an answer was judged”- Find the judgment, by searching for the reviewer, the verdict or words from the answer.
- Expand the row to read Why.
Go back to the answer
Section titled “Go back to the answer”Select the answer’s text in The answer judged. The conversation opens with that answer in view, and with every verdict recorded on it.
Export judgments
Section titled “Export judgments”Admins, Reviewers and Auditors export a tenant’s judgments as a CSV file through the control plane’s API, signed in as themselves (see API overview):
GET /v1/tenants/<tenant-id>/scores/export?from=2026-10-01&to=2026-11-01from and to take a date (YYYY-MM-DD) or a full timestamp. Without from, the export starts on the first day of the current month.
The file has one row per judgment, with these columns: id, created_at, subject_type (assistant_message or gateway_sample), subject_id, source (human or model), name, value_label, value_number, actor, comment, config_digest, trace_id and subject_excerpt. A final TOTAL row gives the row count. If the range holds more rows than one file carries, a TRUNCATED row says so: narrow the range.
config_digest identifies the configuration of the assistant that produced the answer. Two judgments with different digests judge two different configurations; compare like with like. trace_id is empty in a file a Reviewer exports.
Each export is recorded in the audit trail, with its range and row count.
Limits
Section titled “Limits”| Limit | Value |
|---|---|
| Judgments listed | The 200 most recent |
| Rows per export | 200,000 |
| Note | 2,000 characters |
Judgments are kept with the answer they judge. When a conversation or a captured turn is deleted at the end of its retention, or erased, its judgments and notes go with it. The audit trail keeps the fact that a verdict was recorded, without the answer or the note. See Data retention and Erasure.