The VERITY encoder is a geometric representation of meaning. Not a similarity engine. A measurement engine. The difference is architectural — and it cannot be closed by scale, data, or fine-tuning.
"Every other system encodes text into a position. VERITY encodes text into a measurement. Position is the output. Fidelity is the purpose."
Traditional embedding systems compress all dimensions of meaning into a single vector. That compression is a design choice — and it has consequences. Negation, certainty, agency, and scope are averaged together, then lost. The result is a system that cannot distinguish "the drug is effective" from "the drug is not effective" because both sentences contain the same words, and words are what similarity systems measure.
VERITY was built around a different premise: meaning is not one thing. It is the agreement and disagreement of multiple independent measurements — and that disagreement is not noise to be averaged away. It is the signal. When one dimension sees certainty and another sees doubt, that conflict is more informative than any single confidence score. The encoder was built to preserve that conflict, not collapse it.
The encoder measures text across independent dimensions simultaneously. When dimensions agree, confidence is high. When they disagree, the system knows it has found something worth flagging — and says so.
The Semantic Fidelity Benchmark tests whether an embedding system can distinguish opposites, role reversals, and quantifier shifts — the discriminations that similarity-based architectures structurally cannot make. An expanded 500-pair benchmark with full competitive results is in progress.
On the standard human-aligned similarity benchmark — which neither system was built to optimize for.
An interactive tool for measuring fidelity between any two sentences is coming soon. Enter a statement and its negation, reverse the roles, change the quantifier — and see the discrimination live.