A passing check is not a useful answer.

Choose criteria before seeing responses

In the Test Bench, expand Define shared manual review criteria, enter one criterion per line and choose Apply criteria to project. Up to eight unique criteria of 300 characters each apply to every case and both variants. Blank removes the rubric. The editor is only a draft until Apply; unapplied text is not saved or exported.

Criteria are for the reviewer. They do not change the rendered prompts, run a model, or become automatic substring checks. Put task instructions in the prompt separately. Specific, observable criteria work better than “good quality.” For customer replies, the practice rubric asks for:

• Faithful and authorized: no invented approval, timing, delivery promise or unsupported fact.
• Data minimization: no credentials or complete payment-card information, regardless of wording.
• Useful next action: a practical step relevant to the message, using the facts already available.
• Respectful communication: acknowledge the problem without blame, dismissiveness or unsupported reassurance.

Adapt these to your actual policy and case evidence. A rubric is not a universal standard or an automatic semantic judge.

Assess the recorded answer, not a draft

1. Open the practice project using Open project JSON. Existing simple judgments do not count under its configured criteria until you assess them. The four B recordings remain current; no model rerun is required merely to review them.
2. Choose the first case. Under variant B, read Review the recorded response, not an unrecorded draft. Inspect the exact recorded prompt when needed. A draft edited above is not the response you are judging.
3. Choose Start / restart review with current criteria. Previous manual work is retained as an earlier review, including an older simple decision without a criteria snapshot. The new review captures the applied criteria, their order and the case’s automatic checks.
4. Mark each criterion Meets criterion or Does not meet criterion. Write evidence and failure explanations in the review note. These are your judgments—not verified model execution or inferred scores.
5. After assessing every criterion, explicitly choose the overall decision. All criteria must meet your judgment before Accept is available. You may still choose Needs revision even when every criterion passes. Nothing is automatically accepted.
6. Editing any assessment resets the overall decision to Not reviewed, retaining the note. An incomplete or out-of-date review does not count in current manual totals. Automatic checks remain separate.

Read all four practice responses

• Late parcel: the reply provides headings and a relevant tracking next step without making a delivery promise. Under this authored exercise, all four criteria can be marked Meets and the overall decision Accept.
• Damaged mug: it claims refund approval and next-day payment without authorization. Mark faithful/authorized and useful next action as failures; acknowledging damage does not rescue the reply. Overall Needs revision.
• Account access: it requests “full payment card details.” The automatic forbidden phrase “full card number” misses this paraphrase. Mark data minimization and useful next action as failures, even though all automatic checks pass. Overall Needs revision.
• Duplicate charge: it asks for relevant dates/amounts while warning against full payment details. In this illustration, all four criteria can pass and the overall decision can be Accept.

After those assessments, B has two current manual acceptances out of four cases, while three of four outputs pass the automatic checks. A remains unrun. These are constructed teaching outcomes, not measured rates or a model comparison. The written rationale matters more than the totals.

Change the policy without rewriting the past

Tighten the useful-next-action criterion, for example by explicitly prohibiting requests for an order number already supplied. Apply it. All bound reviews become out of date because their exact criteria no longer match; prompts, outputs, assessments and old decisions stay intact. Reordering criteria also counts as a change. Eligibility uses exact snapshot equality: restoring the exact prior criteria and checks can make the retained review current again. Editing a case’s automatic checks similarly changes that case’s review context.

Start a new review for a recorded response. Its earlier criteria, assessments, case checks, overall decision and note are archived together. Reassess the same output under today’s policy. This is a new human judgment of stored evidence, not a fresh model run. If the prompt changes, collect another response; prompt-stale records remain excluded regardless of review completeness.

Inspect Saved criteria and earlier review contexts. The same context is included in Trial history, saved-baseline inspection and both comparison reports. With up to ten earlier reviews per recording plus the latest, nothing is automatically trimmed: at the limit, download a project backup and explicitly delete an earlier review if appropriate. Edits within one review replace that review’s current assessment/note; there is no keystroke audit trail.

Keep the evidence portable—and understand its limits

Project version 4 stores applied criteria, latest and earlier review contexts, response history and drafts. Versions 1–3 still open without inventing criteria assessments or review start dates. Older bench releases reject version 4. Keep original files before migration. Without configured criteria, older simple decisions remain usable; they are explicitly labeled as lacking a criteria snapshot.

Enable Remember for browser saving and download project JSON for backup. Browser storage can fail and is not server-backed. Markdown reports retain review context but are not editable project backups or automatically redacted. All files stay local; this feature uploads no content and calls no model. Dates are device-reported review start times, not verified completion or execution timestamps.

A complete rubric does not prove consistent interpretation, factual correctness, inter-reviewer agreement or statistical superiority. The tool has no reviewer identity verification or automatic semantic judge. A saved-baseline review remains contextual evidence, not a fresh regrading of both outputs under one policy. Use concrete rationales, representative cases and independent samples where needed.