The Truthometer: A Structured Cited Verdict, a Rulebook for the Claims No Register Settles, and a Distilled Judge

Abstract

Paper 5 shows that for the commercial web, truth is a register join rather than a judgement, and that grounding in retrieved evidence is what makes verification work. This paper is the system that follows, and its contribution is threefold: the score, the rulebook, and the distillation boundary. **The Veracity Score.** We summarise a domain with V = F × H, fact integrity times honesty. F is a multiplicative gate over corroborated confirmed falsehoods, so no amount of true boilerplate buys back one confirmed lie; H is the consequence weighted calibration of assertions against retrieved evidence. Regulatory permission is held separate and never folded into V, because a claim does not become less truthful by crossing a border. Every result carries its tier (well formed, key resolves, entity matches) and a corroboration gate: a mismatch is not called a lie without two independent signals. Measured on a gate versus average ablation over 1,248 diverging domains, the gate scores flagged domains at a mean honesty of 0.167 where an average would score 0.326. **The rulebook.** The register join settles identity facts but not "this supplement prevents COVID" or "the election was stolen". The move that reaches those is not to judge each afresh but to recognise when a claim repeats one an authority has already adjudicated: a unified, embedded index of adjudicated claims (its register grounded core the layer the published fact check feeds do not carry), matched by a two stage bi encoder then cross encoder recogniser, returning the authority's own ruling and citation. **The distillation boundary, and it is a negative and a positive.** The judgement the earlier work showed only a language model could perform is a *learnable function*: a compact encoder trained on the regulator's published adjudications reaches an area under the curve of **0.794** on a disjoint held out slice of the rulings, above the **0.703** of the evidence grounded language model it learned from and far above the roughly 0.46 of a text quality instrument, and it needs no language model at the point of use. It generalises across regulators (trained on United Kingdom advertising rulings, it flags 92.1 per cent of United States drug enforcement claims it never saw while false flagging 3.8 per cent of ordinary factual text). The companion negative is the boundary: the claim *extractor* does not distil the same way, a small generative student fabricates, so extraction stays a faithful large model and only the judgement is compressed. This is where distillation works and where it breaks, not a blanket claim that it works. **Scale.** Fifty thousand copies of one claim collapse to one canonical claim, verdicted once against the evidence and cached with its citation, so the per page cost collapses to recognise, normalise and look up; the expensive grounded judgement runs only on genuinely new claims. The recogniser is a property of the page and bakeable; the verdict is a fact about the evidence and is a live join, re verdicted only when the evidence changes. The system returns the record that settles a claim and a pointer to where it lives, never a proprietary opinion. --- ### 10.8 The composition: character, verification, and the sincerity residual The two halves of this paper measure different properties of the same entity. The character instrument measures how a site presents itself: how rigorous, how confident, how commercial its prose is.