Every sentence in a VERITY response is independently measured before inclusion. Sentences that fail verification are excluded. Not softened, not hedged, not qualified. The system would rather say less than say something it cannot prove.
"Language models predict the next word. VERITY predicts whether the next claim is true enough to say. The difference is everything."
Language models generate text by predicting the most statistically likely continuation of what came before. This produces fluent, confident-sounding responses regardless of whether the underlying claim is accurate. The model does not know what it knows. It knows what tends to follow what.
VERITY's decoder works differently. It does not predict words. It retrieves verified sentences and assembles them. Each candidate sentence is measured against the query before inclusion. If the fidelity score falls below the threshold for commitment, the sentence is not included. The system outputs exactly as much as the evidence supports, and no more.
The output of the decoder is not just a response. It is a response with a posture. The posture comes from the evidence for this output, not from frequency statistics. It describes this specific response, for this specific query.
Under Daubert v. Merrell Dow Pharmaceuticals, scientific evidence must satisfy four criteria for court admissibility. Large language models fail all four, structurally. VERITY was designed to satisfy each one.
| Daubert Factor | LLM Systems | VERITY |
|---|---|---|
| Testable & Falsifiable | Non-deterministic. Identical inputs produce different outputs. The methodology cannot be tested because there is no stable methodology to test. | Fidelity measurement is deterministic. Identical inputs produce identical results. The evaluation protocol is specified, stable, and reproducible. |
| Peer Review | Architecture is published. Outputs are not auditable. Attention weights do not explain specific claims in any form subject to meaningful review. | Results are documented. The Semantic Fidelity Benchmark is public. Outcomes are reproducible by any party with access to the same inputs. |
| Known Error Rate | Hallucination rates vary by prompt, domain, and phrasing. The error rate for any specific output is unknown at the time of production. | Every response carries a posture. Error rates are stated per category. Where evidence is strong or weak is available for every output. |
| General Acceptance | Not accepted as a forensic evidence methodology by any professional forensic body. | Measuring whether a claim is supported by verified evidence is established practice, independent of the ongoing controversy over AI generation. |
"An accurate but non-credible system is dangerous. It is right most of the time, so people trust it. But it is equally confident when right and when wrong. The errors are undetectable until harm occurs."
Fidelity as a Measurable Quantity · Credasis