Schema-valid Decisions API calls returned near-uniform, unusable choice confidences

A developer's October 13 public bug report recorded normal HTTP 200 responses from Perplexity's Decisions API while straightforward choice questions returned almost equal alternative probabilities. A strongly negative product review, for example, produced scores close to one third for positive, mixed and negative choices, with very low reported choice confidence. The same tester repeated direct requests and compared versions, then reported that the provider investigated and the behavior later returned to normal.

This supports a concrete first-hand semantic regression report, not an independently measured service-wide outage. Neither the cause nor the population of affected production decisions is known, and there is no documented customer harm. HTTP success and valid JSON only show transport and schema health. They do not prove that a classification, routing or authorization decision is meaningful.

The useful operational question is how a downstream system identifies loss of discrimination while the API remains formally available. Known-answer sentinels, negative controls and task-specific calibration checks can expose regressions that uptime monitors miss. They cannot by themselves authorize a high-impact action or establish which infrastructure layer introduced the abnormal result.