32 / Public failure-case series
These first-party known-answer cases test specific AI-agent and API-service claims. The inputs, expected outcomes and limitations are public so an independent developer can reproduce or challenge them.
01 / Cases
Claim: The agent completed within the agreed 500 ms limit.
Evidence: Measured latency is 781 ms while uptime and freshness remain within their thresholds.
Expected: FAIL — one explicit service promise is breached.
Limitation: The engine evaluates the submitted measurement; it does not independently observe the network request.
Claim: The returned output matches the committed expected result.
Evidence: Availability checks pass, but the observed SHA-256 digest differs from the expected digest.
Expected: FAIL — availability does not rescue an integrity mismatch.
Limitation: A matching digest would prove consistency with the committed bytes, not that the underlying claim is true.
Claim: The agent result is recent enough for the decision.
Evidence: The supplied freshness age exceeds the agreement's maximum age.
Expected: FAIL — a correct but stale result cannot satisfy the current-evidence rule.
Limitation: Clock source, collection method and caller honesty remain external trust assumptions.
02 / Reproduce
Use the documented engine or API against the unchanged inputs and record the runtime, command and timestamp.
Preserve the complete output and limitations. A negative result is accepted when it is reproducible.
Enter the public evidence queue without buying or holding PVTY.
03 / Native demonstration
The existing short demonstration is downloadable for native X and YouTube uploads. Captions must link to a reproducible case and retain the evidence boundary.
Independent challenge