Empirical Evaluation of the Framework for TruthProbability, Perceived Credibility, and Importance Scoring

This empirical study evaluates a unified framework for scoring statements along three distinct outputs: evidence-weighted truth-probability T_abs(s), receiver-dependent perceived truth-probability T_perc(s,r), and receiver-dependent importance I(s,r). The study uses a blind external-scoring design in which GPT-5.5 Thinking assigns atomic framework variables rather than final benchmark labels. Final scores are then computed by the mathematical formula engine and joined to hidden targets only after scoring is complete. Three datasets are used because each corresponds to a different output type: HumAID for crisis-monitoring importance, CREDBANK/CredEvent TurkRatings for perceived credibility, and LIAR/PolitiFact for evidenceweighted truth. Across the three tests, the framework produces moderate-to-strong empirical alignment with benchmark targets. The strongest evidence is in discriminative and ranking performance, with binary AUC values in the acceptable-to-excellent range and correlations reaching large-effect territory on the harder credibility and factuality tasks. The results support the framework as a structured scoring architecture while identifying calibration issues that remain most visible at distributional extremes, especially in graded factuality scoring.