Every sphere in Tin-Foil ends on one word — REAL, FLUFF, or DEBUNKED. This is the machine that word comes from. An assay weighs the evidence for a claim against how extraordinary the claim is, in log-odds — the exact meaning of ‘extraordinary claims require extraordinary evidence’. Tap the scale: LEFT to add weight against, RIGHT to add weight for. Or load a real case and watch where it settles.
The needle is Bayes' theorem in log-odds. Start from a prior (how ordinary or extraordinary the claim is, as odds), then add each piece of evidence as its weight of evidence — log₁₀ of the likelihood ratio (how much more expected the clue is if the claim is true than if false). Posterior log-odds = prior + Σ weights; this is exact and additive (I.J. Good's decibans, Turing's team at Bletchley Park, 1941). One ban = 10× the odds. The verdict bands are honest cutoffs: past +1 ban the evidence decisively supports (REAL); past −1 it decisively refutes (DEBUNKED); in between, nothing has moved it — unfalsifiable or unsupported (FLUFF). Node-checked: the three worked cases land on the same verdicts the domain already gave them (Illuminati → DEBUNKED, life-elsewhere → FLUFF/open, the Cathedral → REAL).
The arithmetic is exact; the inputs are judgment. How extraordinary is the claim? How strong is a blurry photo, a leaked memo, a witness? Those numbers are estimates, and a dishonest weigher can rig any verdict by inflating a clue's likelihood ratio — garbage in, garbage out. The instrument shows the structure of honest weighing, not an oracle that reads the world for you; it makes the assumptions visible and additive so they can be argued, not a machine that ends the argument. The worked-case weights are our reasoned estimates from the cited record, not measured constants.