A request for medium items in red or blue should not return every red item, every blue item and every medium item. Under the custom search contract tested here, the colour choices form one group, joined by OR. That group is joined to the size constraint by AND. In the fictional catalogue below, the exact answer is I03 followed by I05.
A retry is justified only when the next execution has a plausible path to a different result and the operation is safe to repeat. A malformed shipment reference needs correction. An expired session needs authentication re-established through the application. A temporarily unavailable status service may justify another attempt within a limit. Treating all three as "try again" hides the condition that must change.
Choose an output change by testing the consumers that must keep working, not by classifying the change as additive or assigning it a new version number. In the synthetic contract tests below, an added field worked for an old consumer that permitted extra keys and failed for an old consumer that required an exact key set. A renamed field needed an explicit adapter for either old consumer. An unchanged schema still carried the wrong payable total.
Split a broad tool when its action parameter changes what the caller is allowed to do, what state can change, or what a successful result means. A tool that checks availability, creates a booking request and commits a booking covers three different promises. Separate contracts make those promises explicit. They do not, by themselves, enforce authorization or prove that an AI assistant will select the right operation.
Use an explicit currency code and an exact amount representation, then define the currency scale and the point at which rounding occurs. For a quote tool that accepts sub-minor-unit prices, decimal strings can preserve the input until the total is calculated. An integer minor-unit field is a better fit when every accepted amount is already settled to that currency’s minor unit. Neither representation supplies the missing business rules by itself.
Accept a declared timestamp format with an explicit UTC offset, reject locale-dependent date strings, and treat named-timezone checks as a separate requirement. A value can identify a valid instant yet still be wrong for a booking's local service day. The tool needs to distinguish those outcomes before it changes a schedule.
Return alternatives when the task is to compare equally valid matches. Request another input when the task requires one result and a missing, meaningful requirement would distinguish the candidates. Sorting tied records by a stable identifier makes the response repeatable. It does not make the first record more suitable.
A search tool can promise that returned items match stated filters. A recommendation tool must also explain the basis for preferring an item for a purpose. Returning the cheapest matching product answers a price preference; it does not, by itself, establish that the product is suitable for the job.
An audit that cannot reach an order, account or confirmation state should report blocked coverage, not silently score the task as a failure or a success. Name the missing prerequisite, record the last state actually reached and leave the task outcome unknown. A failure finding requires a task attempt and an observation that contradicts its expected outcome.
A failed audit task warrants a defect fix when the application promises that task, the required conditions have been established, and the observed result violates the promise. A feature deliberately left out is a different finding. It may warrant new development, but a failed attempt to call it does not prove an existing implementation is broken.