This is a 2D-native reading of the same Bayesian fusion math as the 3D orbiting-evidence lab: instead of watching evidence nodes circle a claim sphere, you watch a log-odds waterfall β each active source's weighted contribution stacks left-to-right into a running total β next to a sigmoid inspector curve that shows exactly where that running total lands on the S-curve mapping log-odds to probability.
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 β [-1, +1] (support β refute), degraded by the noise slider. Evidence combines additively in log-odds space β the standard Bayesian fusion rule for independent evidence β then converts back to a probability with the logistic (sigmoid) function:
L = Ξ£ w_i Β· s_i Β· active_i Β· (1 β noise)
P(claim true) = Ο(L) = 1 / (1 + e^(βL))
- Waterfall bars (top) β each usable source's signed contribution wiΒ·siΒ·(1βnoise), stacked in source order; the step line traces the cumulative log-odds L after each bar.
- Sigmoid curve (bottom) β the fixed Ο(L) function plotted over L β [-6, 6]; a marker dot rides the curve at the current L, reading off P directly as a height on the S-curve.
- Reliability threshold β auto-drops any source whose weight falls below the cutoff, greying its bar out.
- Data-quality noise β shrinks every included signal toward zero, shortening its bar and pulling L toward 0 (P toward 50%).
- New Claim β draws a fresh hidden ground truth and resamples every detector's signal around it.
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 β this is a defensive verification model, not a tool for generating disinformation.