Calibration

Accuracy & FPR — published, not claimed

We commit to a per-use-case false-positive rate, abstain when we can't meet it, and publish the number here. Targets below are engineering constraints; the calibration state is live engine output. We never show a measured figure the engine can't back.

Per-use-case false-positive rate

Use case	Target FPR	Measured FPR	Detection recall	State
Consumer / Trust & Safety Image + video calibrated; audio abstains.	< 5%	pending live feed	99.75% (image)	Calibrated
HR / Recruitment Interview-fraud profile; audio abstains.	< 2%	pending live feed	98.99% (video)	Calibrated
Banking / Identity Locked pending real enrollment data — we abstain rather than assert.	< 1%	pending live feed	—	In calibration

Measured FPR publishes from the engine's calibration endpoint (contract-first; wiring pending). Until it is live, this column reads “pending live feed” — by design, it is never a placeholder number.

Validated detection accuracy

0.9999

Image detection AUC

99.75%

Image recall @0.5% FPR

98.99%

Video recall (v7)

Abstains

Audio — in calibration

Generalization, proven

A detector that only works on the generators it trained on is worthless against tomorrow's. Our DFDC-only model scored 0.62 AUC — a coin flip — on an unseen generator. The v7 backbone (Unidata fine-tune) fixed that blind spot to 0.99 AUC. We re-test against new generators and recalibrate through the governed flywheel — generalization is monitored, not assumed.

Why “published, not claimed”

Every decision is stamped with the engine and policy version that produced it and recorded in an immutable audit trail. When a use case isn't calibrated to its target, the engine returns “inconclusive” with a recommended next step — it does not guess. That is what lets us put an error rate on the page instead of a marketing claim.