Real OSINT verification never relies on one signal — it fuses several independent, individually unreliable detectors into one confidence estimate, the same way a spam filter or a medical test combines evidence. Each source here is modeled as a noisy detector of a hidden ground truth, with a fixed reliability weight wi reflecting how diagnostic that signal type really is:
Cross-source corroboration w = 1.1 (strongest)
Geolocation / EXIF match w = 0.9
Domain WHOIS age/reputation w = 0.7
Archive.org snapshot diff w = 0.6
Author credential check w = 0.5 (weakest)
Each active source reports a signal si ∈ [-1, +1] (support → refute), degraded by the noise slider. Evidence combines additively in log-odds space — the standard Bayesian fusion rule for independent evidence — then converts back to a probability with the logistic (sigmoid) function:
L = Σ w_i · s_i · active_i · (1 − noise)
P(claim true) = σ(L) = 1 / (1 + e^(−L))
- Evidence toggles — include/exclude a source from the fusion, exactly like an analyst deciding which checks are available for a given claim.
- Reliability threshold — auto-drops any source whose weight falls below the cutoff, simulating a stricter evidentiary bar.
- Data-quality noise — shrinks every included signal toward zero, modeling degraded, incomplete, or tampered source material.
- New Claim — draws a fresh hidden ground truth and resamples every detector's signal around it, the way a new viral post presents a fresh evidence trail.
Verdict bands (P > 70% likely true, P < 30% likely false, otherwise uncertain) mirror the confidence tiers real fact-checking desks use before publishing a rating — this is a defensive verification model, not a tool for generating disinformation.