The Tokenizer Never Lies
While API gateways can spoof headers, model IDs, and response formatting, they cannot forge the physical byte-pair encoding of the backend model. Comparing observed token counts against calibrated ground truth exposes the underlying architecture.
Relative Error Scoring
Absolute token counts bias detection toward models with smaller tokenizers. By normalizing errors relative to baseline targets across multiple calibrated probes, Unmask AI achieves mathematical separation between model families.
Empirical Verification
We test conformance, tokenizer mechanics, behavioral signatures, and cognitive horizons to produce an immutable verdict with cryptographic audit trails.