These four properties are rarely offered together.
Four properties that voice-AI products rarely offer together — and the reason you can act on a KKOS model instead of guessing at it.
Six explicit consents before a session. Raw media is not stored. Keys stay server-side. Delete your model and all derived data goes with it.
Every trait arrives with its evidence, a confidence value, an uncertainty statement and alternative readings.
Mark any trait right, wrong or not sure. Your feedback re-weights the model instead of being quietly ignored.
Structured JSON, versioned models, same input yields the same output — so results can be compared and audited over time.
Two are required to build and save a model.
Every trust property on this page is backed by a published measurement. Three calibration reports cover the engine itself — its repeatability, its honesty about what it does not know, and the signals it actually receives — each stating the challenge, the approach, the result and what it implies, together with the exact data file and command needed to reproduce it.
One fixed conversation, modelled three times independently. Stability score 0.7333; the worst dimension swings 55 points between runs.
Read the report REPORT 0274 real conversations produced 35 models. Trait coverage is 100%, yet 40% of those models return an explicit insufficient-data flag instead of generating a guess.
Read the report REPORT 03Text arrives in 21 of 21 samples, voice prosody in 12, face signal in 1. We publish the gap rather than advertise the capability.
Read the reportReports are published only at the coverage level the data supports. Where coverage is near zero — face and expression signals — this site does not claim the capability.