A form can pass its demonstration while saving an event that ends before it starts. Enter a sensible pair of times, press Save, and the success message looks convincing. Ask a second assistant whether the implementation satisfies the request, and you may get another explanation of the same successful path. The missing evidence is what happens when the relationship between the fields is wrong.
Give quality assurance a different question: which acceptance claim can you disprove on this build? That question needs a requirement, an identifiable artifact, and a place to preserve observations. Asking for skepticism without those inputs can produce an impressive list of imagined defects. The useful result is a violation another person can reproduce.
Consider a synthetic event form for a fictional workshop. Its agreed scope is same-day events, entered as local wall-clock times, with an end strictly later than the start. Overnight events are outside this example. Invalid input should leave the form available for correction, explain the ordering problem, and create no saved event. These are chosen requirements for the example, not universal rules for scheduling software.
The implementation handoff says that both time fields are required and the form saves valid events. Its demonstration uses a start of 09:00 and an end of 10:00. Those values establish one useful path if the saved record is checked. They do not establish that the application compares the fields. Two individually plausible times can still describe an invalid event under the agreed rule.
That distinction has a technical basis. The browser provides basic field constraints, while more complex relationships can require additional validation. MDN’s constraint validation guide describes combining constraints across fields. The documentation supplies a mechanism, not evidence that this fictional form implements it. The reviewer still needs to exercise the claimed behavior.
Write the QA brief before forwarding the implementation’s explanation. State the requirement, the exact build identifier, the permitted test environment, and the acceptance evidence already available. Ask the reviewer to attempt a small set of counterexamples and return expected behavior, actual behavior, and reproduction steps for each violation. Require a separate list for unanswered questions. Keep code changes outside this review assignment.
For the worked packet, call the fictional build demo-A and use an empty disposable event store. Open the form, enter a start of 10:00 and an end of 09:00, then press Save. Suppose the form announces success and the stored event retains those values. That is the hypothetical observation being examined here. No application was executed for this article, and demo-A is an illustrative identifier rather than a released build.
The corresponding finding is specific: demo-A accepts an end time earlier than its start despite the same-day ordering requirement. Its expected result is an ordering error with no new record. Its actual result, in the hypothetical case, is success plus a saved invalid record. The evidence should include the exact input, the action sequence, the visible response, and the saved-state check. A screenshot of the success message alone cannot establish that the write occurred.
That finding belongs in the blocking category once it reproduces on the specified artifact. The reason is the violated acceptance rule, not the reviewer’s tone or the size of the patch needed to fix it. If the reviewer cannot access the saved state, the finding must say so. They may have evidence of misleading feedback, while persistence remains unverified. Those are different claims and should remain distinguishable.
A neighboring input sharpens the requirement. Set both fields to 10:00. Under this example’s strictly-later rule, equality must fail too. If the brief had said only that events cannot end before they start, equality would remain a question for the owner. The reviewer cannot silently convert an ambiguous sentence into a new product policy and then call the implementation defective for violating it.
Keep a normal event in the packet as well. A 09:00 start and 10:00 end should save exactly one record with the entered values. Reset the disposable store between cases so an earlier record cannot masquerade as the current result. Playwright’s testing guidance recommends testing user-visible behavior and isolating tests from one another. Those principles support this proposed review design; they do not prove that any particular test suite covers it.
The small packet now has three time pairs, explicit outcomes, and a shared rule. The reviewer can challenge the implementation without guessing what success means. A real run would add the actual build revision, browser and environment details, evidence locations, and observed outcomes. Until that run happens, the observed-result fields stay empty. A filled example must remain labeled as an example when someone copies the template.
There is a temptation to make the reviewer repair each problem immediately. That collapses two useful responsibilities. While reviewing, the assistant should preserve what failed and why it matters. A repair changes the artifact under examination and needs a new identifier before retesting. Otherwise a report can quietly mix a failure from one version with a passing observation from another.
There is also a limit to what a different prompt buys. A second pass can carry the same mistaken assumptions as the first, particularly when both begin with the implementer’s explanation. Separating the requirement and asking for counterexamples gives the review a different starting point. It does not establish statistical independence or guarantee defect detection. The reproducible finding, checked against the agreed requirement, is the evidence worth carrying forward.
Nor does a browser rejection establish the whole storage boundary. MDN explicitly warns that client-side validation does not replace server-side validation. If the acceptance claim includes protecting stored records against direct requests, the agreed review scope needs evidence from that boundary too. If access permits only the form, record that limit instead of expanding permissions or declaring the entire application safe.
The owner receives a compact decision packet: a confirmed requirement violation, any unresolved policy questions, the cases actually checked, and the boundaries left untested. They can send the reproducible defect for a bounded repair or resolve an ambiguity before work continues. Optional interface preferences belong elsewhere unless they obstruct the accepted task. QA supplies evidence for that decision; it does not acquire authority to redefine the task.
Before accepting the form, require someone to replay the reversed-time case on the exact candidate build and inspect the resulting state. A second approving paragraph cannot do that work.
Use the Playwright testing guidance to check isolation and observable outcomes in your next QA brief.
If this was useful, forward it to one engineer who needs less noise in their feed.


