Inconclusive results
A binary pass/fail system has to guess when the evidence it needs is missing. That guess is usually an accusation against an agent that did nothing wrong.
DialogAssert has a third outcome.
PASS the control held, and here is the evidenceFAIL the control did not hold, and here is the evidenceINCONCLUSIVE the control could not be evaluated in this environmentWhat inconclusive means
Section titled “What inconclusive means”INCONCLUSIVEUnable to verify selected card because structured tool argumentswere not available for this run.
This is an assurance gap, not an agent failure.Inconclusive controls are excluded from the pass rate entirely rather than counted against the agent. They are reported separately, and they are excluded from the release gate — a control nobody could evaluate should not silently approve a release either.
Assurance coverage
Section titled “Assurance coverage”Coverage answers a different question from correctness:
- Behavioural correctness — what did the agent get wrong?
- Assurance completeness — what are we currently capable of proving?
Release candidate v1.8.4
Approved controls 23Deterministically verified 15Semantically evaluated 5Not verifiable in this environment 3
Assurance coverage 87%Critical controls covered 100%
Result BLOCK RELEASEReason 2 new High regressions 1 unstable control 3 evidence gapsThe second number is often the more useful one to an AI platform or risk team. It says: your testing standard requires evidence of approval state, and this agent’s instrumentation cannot provide it. That is a control-gap finding about the environment, and it is usually fixable within a sprint.
Why this is not a weakness
Section titled “Why this is not a weakness”Reporting what you could not verify is what separates an assurance system from a scorer. A report claiming 94% pass with silent gaps is less trustworthy than one claiming 87% coverage and naming the missing 13%.