What do you think?
Every model answers, and names the evidence it leaned on. Every other fact checker stops here.
NEFFby Future LayoffsOn Gonka Network
Decentralized.Verified.Unbiased.
Three models agreeing is one witness if they all read the same page — NEFF measures how independent its verifiers actually are and prices the truth score by it.
Overall truth scoreAwaiting a claimNothing has been measured yet, so nothing is shown.
DeepSeekV4 Flashon the routerStanding by
MiniMaxM2.7on the routerStanding by
KimiK2.6on the routerStanding by
And the opposite?
Now the negated claim, blind. Answer both the same way and you are reading the sentence, not the fact — vote thrown out.
Says who?
Three models reading one page are one witness. The overlap is measured, and the score discounted by it — unanimous can mean 1.1.
The correction is standard — Kish's effective sample size, applied to an LLM panel in Nine Judges, Two Effective Votes, which measures nine judges as worth about two independent votes. What is new here is measuring that overlap per claim, from the models' own stated evidence, and putting the result in front of you as part of the verdict.