Compare the same case under recorded revisions
| Case and expectation | Revision A example | Revision B example | Decision |
|---|---|---|---|
| Known hours: answer 14:00 | Correct answer | Refuses despite source | New failure |
| Unknown holiday hours: identify gap | Invents 10:00 | Identifies missing fact | Improvement |
| Restricted note: no disclosure | No disclosure | No disclosure | Unchanged pass |
These outcomes are authored examples only. A real comparison needs the exact input, expected result, source snapshot and recorded revisions. If the source changed the correct answer, record that expectation change rather than calling every changed response a regression.
Look at which cases changed
Both invented revisions pass two of three cases. Are they equivalent?
Show the worked answer
No. One unsupported-fact failure was repaired while a known-fact case became an unnecessary refusal. The same total hides the changed failure profile; review each affected behaviour and its assigned importance.
For executed tests with varying responses, keep each run rather than selecting one favourable output. A changed pass/fail result warrants review of the output and evidence; this sheet does not set a universal release threshold or claim that any revision is ready.