Is this answer reliable? The output is a single reliability quotient on the final answer. Everything else, corroboration, sourcing, bias, relevance, is the process that produces that one number. Reliability rides on the answer, not on the pedigree of the process.
Many fact-checkers decide truth by consensus among groups they consider credible. That is itself a bias: it is appeal to authority plus groupthink, and choosing who counts as credible is where the bias enters. Agreement is one signal, not a verdict. A claim is reliable when it also survives contradictory evidence and traces to evidence rather than to an authority's stamp. COMING is the contradiction/source axis that carries this; BUILT today is the corroboration meter, whose limit is the one below.
Ask the model the same question several times, so it produces a panel of independent answers. Each answer is placed as a direction in meaning-space, and the engine measures how completely those directions collapse toward a single point. Tight collapse means the answers corroborate each other on the detail that matters; a scatter means they do not. It reads the question to know the decisive detail: a quantity (collapse the numbers), a yes/no or verification (collapse the verdict, with canon counting as fact within a fictional work), or open prose (collapse the shared content). The result is a 0 to 100 quotient shown with the panel, so you can see why.
BUILT When the model fills a blank, the tool never passes the guess off as certain. It rates the fill-in (the quotient), and when corroboration is low it surfaces the divergent answers as alternate interpretations, none corroborated, rather than picking one silently. The model stating its own justification for a fill-in is COMING.
BUILT It abstains sparingly, and abstaining is never censorship: it still shows the answer. It declines to vouch only when the question itself is a hallucination trap, an invented entity or an unknowable specific, where a model is most likely to fill the blank with something confident and wrong. Everyday answerable questions are not abstained.