The claim starts with a prior probability P(H) you set directly. Every evidence node carries a likelihood ratio LR = P(E|H) / P(E|¬H) — how many times more likely that evidence is to appear if the claim is true versus false. Toggling a node TRUE applies its LR; toggling it FALSE applies 1/LR (the evidence's absence is itself informative); UNKNOWN nodes are skipped entirely. Because probabilities are awkward to chain, the update runs in log-odds space, exactly equivalent to repeated Bayes' theorem:
odds(H) = P(H) / (1 − P(H))
for each observed evidence i:
odds(H) ×= LRi (if TRUE)
odds(H) ×= 1 / LRi (if FALSE)
P(H | evidence) = odds(H) / (1 + odds(H))
This is naive-Bayes sequential updating: each piece of evidence is assumed conditionally independent given the claim, so its contribution to the log-odds is simply ln(LR) added or subtracted. A LR far from 1× (either direction) swings belief hard; a LR near 1× barely moves it — sliders let you feel that directly.
- Belief meter — the claim node's fill and the bar both track the live posterior P(H | evidence).
- Edges — green pulses flow inward for supporting evidence in play, red for opposing, dashed gray for evidence left at UNKNOWN (not used in the update).
- Edge thickness — proportional to |ln(LR)|, i.e. how strongly that node can move belief when active.