Verification
Support is checked. Numbers are gated. There is no score to hide behind.
Every generated sentence is traceable to the passages offered at write time. Checking those passages is a separate step — and a conflicting figure wins over the model’s opinion.
The negation problem
Consider a claim that “the approach temperature was within guarantee,” and a passage stating it exceeded guarantee. The vectors are nearly identical. The meaning is opposite.
That is why cosine similarity is a diagnostic on the evidence panel, never the verdict. Nearness in embedding space is not entailment.
The numeric gate
Labelled quantities are extracted from the claim and from each cited passage. Units are normalised. Values are compared within a configurable relative tolerance to absorb rounding in prose.
Same label, same unit, value outside tolerance →contradicted, regardless of what the language model concluded. A figure that appears in no cited passage cannot be markedsupported.
Years, section numbers, and figure numbers are excluded from the gate. The tolerance is a parameter, not a constant — scientific readers will ask, and they should.
Why there is no confidence score
A blended float implies a calibrated probability that does not exist. Reviewers cannot act on 0.73. They can act on unsupported, contradicted, or partially supported — especially when a cited passage came from a superseded document version, which caps the band at partially supported.
The six verdicts
- Pending — verification has not finished.
- Not a claim — connective or transitional; nothing checkable.
- Supported — cited passages carry the claim.
- Partially supported — partial carry, or support only from superseded versions.
- Unsupported — the passages do not carry the claim.
- Contradicted — a cited passage conflicts, including numeric conflict.
When checking is unavailable
If the checking model cannot run, verdicts staypending. They never default to supported.
Honest limits
Entailment checking is model-assisted and imperfect. The deterministic guarantees are the numeric gate and the export block: client-mode export refuses while any claim is contradicted or any section remains a placeholder, listing the specific reasons.
Review-mode export carries citations as footnotes naming the source document and version, plus an appended verification summary.