Keep The Same Small Test
A fluent summary can still change a commitment, reuse a corrected number or hide a disagreement. This reference pack gives you three small source sets whose facts you can inspect yourself. Keep them unchanged when comparing an earlier run with a later one.
All sources are invented. This is an editorial practice pack, not a benchmark, vendor test result or proof that a service is reliable.
Get The Pack
1. Test input: the shared instruction and three source sets. Send one case at a time in a fresh conversation, using the same instruction each time.
2. Answer key: facts and failure conditions for your review. Keep this out of the model’s input so it does not supply the answers you want to test.
3. Comparison worksheet: a place to record two runs, differences and the next question. The blank file contains no saved tools or private browser notes.
What Each Case Checks
| Case | Source problem | What to watch for |
|---|---|---|
| S-01 | A possible review date and an unassigned approval | A proposal becoming a promise |
| S-02 | A count explicitly corrected by a later note | Reusing or adding the superseded number |
| S-03 | Two estimates with no agreed replacement | Picking one as final or inventing a compromise |
In S-02, the later note explicitly replaces the earlier count. In S-03, neither estimate is approved or marked as a replacement. “Use the newest message” is not a safe rule for both cases. Preserve the stated relationship between the sources, not just their order.
Record The First Output Before Correcting It
Keep the exact prompt and output in an approved local location. Write down the date, tool and any visible account or model settings; use “not shown” when a setting is unavailable. Do not guess hidden model versions or upgrade an account merely to complete the pack.
Compare claims with the answer key rather than requiring identical wording. Mark each case meets these checks, needs correction or could not test. If you use a repair prompt, keep that output separate from the first attempt. A corrected result is not evidence that the initial run passed.
Do not combine the three outcomes into a tool score or ranking. A small, hand-picked set does not represent all documents or future results.
Return When There Is A Reason To Compare
On a later visit, rerun the unchanged inputs if a relevant product change, instruction change or workflow question justifies it. Keep the original worksheet and record the new run separately. If you change both the input and the settings, note that you cannot attribute differences to one cause.
Use your shortlist for a brief, non-sensitive reminder of the observed failure and next check. Keep full outputs and the worksheet outside the short note. Your browser-only checklist is not an evidence archive.
For a different task, adapt a copy of the pack and write its expected facts before testing. Do not replace these invented sources with confidential material in an unapproved service. A successful practice run is not approval for sensitive or high-stakes work.
Continue with the retest workflow or record a completed check without hiding a failed result.