PERSPECTIVE / AI reliability
When is an action complete?
The relationship between generating answers, executing work and verifying outcomes.
The evidence of completion must be defined within the task and its acceptance criteria.
Answer, action and outcome are distinct stages.
A system may produce the right answer without doing the work. A tool call also does not prove the expected result occurred. Completion must be assessed through the outcome in the authoritative system.
Verification is part of the design.
A task's expected outcome must be defined at the outset. It should be clear which record will change, which check will run and who decides in an exception.
Learning through evidence.
Verified outcomes, failures and exceptions inform future system improvements. Evaluating only successful examples hides the real boundaries.
Completion criteria are defined before the task.
Updating a record, preparing a report and coordinating operations require different evidence. That evidence should not be determined by interpretation after a task is complete.
First, clarify the expected output, source system, verification method and acceptance owner. This creates a measurable link between action and success.
Example: completing a record update.
A successful tool response is an initial signal. It does not, by itself, show that the correct field in the target record changed to the correct value under appropriate authorization.
In this conceptual scenario, the result is checked again against the target system's record. If the expected change did not occur, the task should not be marked complete; the exception path should run.
A negative result is also system knowledge.
Verification is not designed merely to produce a positive signal. It must expose partial completion, conflicting records and unexpected behavior.
These records can inform the next evaluation set, architectural change or working rule. Documenting boundaries helps us understand which work the system can handle.


Talk to us