ProjectionBench asks whether models put feelings in your mouth

Field notes

The benchmark evaluates several kinds of affect attribution, including claims made after an explicit prohibition and assertions of hostility or tone. Its calibration scenario deliberately includes a user who genuinely expresses frustration, so the scoring is not designed to reward silence alone. The page also exposes versioned rubrics, raw metric caveats, and transcript-level findings, while noting that API and web surfaces should not be compared.