How every number on this site is computed

The one formula

Row contribution sign × (2 × Truth − 1) × Confidence × Link × Imp × Uniq

Positive contributions are added into POS and negative ones into NEG as magnitudes, so a row lands on the side its sign puts it on and not the side it was filed on: a refuted objection counts as support. The belief score is POS − NEG, open ended. It is a quantity of argued weight, not a probability, and comparing it between a page with forty rows and a page with four is comparing volumes.

Every scored row on every page, whether it is an argument, a cited finding or a prediction, contributes this and nothing else.

Where a page starts, and why it has to start somewhereEvidence →

A row contributes (2 × Truth − 1). A page with no rows reads 0.50, so it contributes nothing, so every page above it also reads 0.50. By induction every page of an argument graph in which nothing is cited reads exactly 0.50, forever, however many reasons are listed and however well they are argued. Reasoning about reasoning never touches the world. What touches the world is evidence, so a page may say what it rests on, and that sets where its truth starts before any of its own rows count.

Starting point p₀ = 0.5 + 0.5 × ESIW × (2 × ERP/100 − 1)  ·  Weight w = k × 2 × ERQ / (ERQ + 1)

Truth, argued (POS + w × p₀) / (POS + NEG + w), k = 1

And then the conclusion cannot outrun its premises

What a page’s rows argue is not the last word on it. A page may list components it needs to be true, and mark some of them load-bearing: a premise the claim does not survive without. A conclusion cannot be more settled than something it needs, so the published truth is the smaller of what the rows argue and the weakest of those premises.

Truth, published min(Truth argued, the smallest truth among load-bearing components that have a page of their own)

It is a minimum and not a replacement, so a premise better established than the conclusion leaves the conclusion where its rows put it, and a premise stated in words that nobody has opened a page for caps nothing rather than capping at zero. This rule is why the five beliefs here read 0.50 while citing twenty findings between them. Their rows argue higher; each is held down by a necessary premise nobody has argued, which sits at 0.50 because nothing beneath it is established either way. That is the rule working, and it is also the thing to check first when a number looks wrong: each belief page says in words what is holding it and names the premise.

ERP is the share of independent replications that agreed and decides which way a claim is pushed; ESIW is the source type and decides how far it can be pushed; ERQ is how many replications there are and decides how much arguing it takes to shift. A page that names none of them reads p₀ = 0.50 and w = k, which is the neutral vote the rule always had, so citing nothing is not a penalty and is worth nothing. The weight is capped at twice k, so no pile of citations puts a claim beyond argument: at k = 1 a handful of argued objections pulls the best-sourced claim back under the line.

Two kinds of page score themselves differently

A linkage, uniqueness, equivalence or driver page is argued like any other claim and scores by the rules above. Two are not.

Importance page max over the interests it lists of Truth(the interest) × Bears(how much the row speaks to it)

A maximum and not a sum, because a reason that speaks to one interest that matters is important; listing five interests it barely touches is not five times as important. Bears is itself an argued page where one exists, and the labelled constant for relevance where none does. An importance page with nothing listed under it reads its own starting point, which is 0.50 unless it cites something.

Media page quality from its own argument rows, impact from its impact rows, each by the argued-truth rule above

A work is scored on whether what it says holds up and, separately, on how much of the corpus it moved. The two are kept apart because a widely read work that is wrong and an unread work that is right are different problems.

The evidence tiers

The sixteen ranked rows are the wiki’s own table with its weights and its order. Two readings an auditor should have: the wiki gives one band, 0.85 to 0.80, for randomised trials and meta-analyses together, split here in the order it lists them; and the wiki has no row for a primary official record, so one was added and marked, because calling a Supreme Court holding a historical trend would be wrong. Classify what the sentence on the page asserts, not the instrument that produced it: “58 percent told a pollster X” is a published statistic, while “Americans believe X” inferred from the same poll is survey evidence and much weaker.

This is the part that is easy to skip and is load-bearing.

Confidence: how much a score has earned the right to count

Components that do not apply to a page are dropped and the rest renormalised, so a page is never punished for a signal its shape cannot carry. Across these 2173 pages confidence runs from 0.00 to 0.67, mean 0.06. The wiki also lists behavioural signals: up and down votes, weekly visitors, dwell time, edit frequency, duplicate submission attempts, per-argument evaluation responses, and the standard deviation of a score over time. A published corpus has one snapshot and no users, so all seven carry weight zero and are reported as having no data rather than scored as zero. Wire them up and they take weight from the structural components.

A score means nothing until the work behind it has been done, and then means more and more as it is done. Confidence multiplies what a page passes to any page above it: at 0 a claim moves its parent not at all, however true it looks. It is deliberately not a cap on the page’s own truth score. Truth is what the arguments say; confidence is how much to bet on it.

What would change the answer

The sweep reaches 6 levels below a page, and every page says so along with how many pages sit deeper than that. The flip point, where a page crosses 0.50 and the conclusion changes sides, is found by bisection to 16 places rather than by algebra, because a closed form would be a second implementation of the rule and would drift from the first. An input that moves a page by less than 0.005 is reported as inert: nothing anyone could learn about it changes the answer. A second sweep holds the input’s confidence at 1, which is what it would be worth once the work behind it is finished, and the gap between the two sweeps is the value of doing that work.

What it does not do. This is one input at a time. It finds single points of failure and it cannot see three assumptions each moving a little in the same direction, which is how correlated assumptions actually fail. The three widest inputs are also pushed against the belief together, which is a gesture at the problem rather than a solution to it. Reading a robust column here as “the conclusion is safe” is the mistake this paragraph exists to prevent.

On every belief and claim page, each claim beneath it is held at false and then at true, one at a time, and the whole graph is recomputed.

ReasonRank: how much depends on a claim

A walk starts evenly at the 162 beliefs and steps from each page to the pages it reads, choosing among rows in proportion to Link × Imp × Uniq, with a 0.85 chance of stepping on rather than restarting at a belief. ReasonRank is the share of that walk arriving at a page; it sums to 1 across the corpus and converged here to a residual of 8.8e-13. The truth of the claim being ranked is deliberately absent, so a page does not drop out of the queue at the moment it is proved false, and confidence is absent, because a page nobody has started is exactly what the ranking exists to find. Work value is ReasonRank × (1 − confidence): high rank with the work done is a settled foundation, high rank with the work undone is the next week an analyst should spend.

The network half of the algorithm the project is named for, which is a different question from whether a claim is true.

Computed equivalency, and what it gets wrong

Two lexical signals over the claim text, each weighted by inverse document frequency across this corpus and averaged: overlap of content words, and cosine over character 4-grams. Pairs above 0.45 are flagged for a human to read; above 0.70 they are probably one page. It measures wording, not meaning. It will flag two claims that share a long subject phrase and say different things, and it will miss two claims that share no words and say the same thing.

It changes no score and cannot. The wiki blends the computed and argued halves with weights set by a validity comparison argument. No such argument exists, so the computed weight is 0 and the argued page decides, which is the same rule as everywhere else here: an input nobody has argued counts nothing. The same goes for the uniqueness multiplier. The wiki defines it as one minus the highest similarity to any other row, computed; this site leaves it presumed distinct and flags the overlap instead, because a number that silently discounts an argument and has no page behind it cannot be audited, and every other number on this site is a link to the page that argues it.

40 pairs flagged out of 2041 claims, of which the 12 most alike are shown and 28 are not. 30 sit on the same page with no uniqueness page between them, which is the live padding risk: each is scored as if it made a point the other did not. A word carried by almost every claim is skipped when looking for pairs, because the score is weighted by how rare a word is and such a word carries no weight in it; a word that is merely common is not skipped. 337,302 pairs share at least one such word and 31,956 of them were scored in full, which covered every claim.

The wiki splits the equivalency score into a computed half and an argued half. The argued half is the equivalence page. This is the computed one.

The labelled constants

What the engine reads when nobody has argued a factor. Each is a presumption, and each is visible in grey on the page so a reader can see which factors are still resting on one.

What this cannot do

A methodology without this section is marketing.

The tables behind every page

• The two tables and the topics, in every shape: JSON, XML, SQL schema, SQL data or a SQLite database.
• Every page's computed numbers, indexed, and every claim in the tables, drafts included.
• Every draft and what it still needs.
• What changed since the last revision, row by row.

Everything the site is built from, free to download and check. Every number on every page can be recomputed from these.

Where the code is

Every rule above is one short module, and a conformance corpus of twenty-four pages, their expected numbers, the five constants and the evidence tiers, published as conformance_corpus.json and conformance_expected.json so an implementation in any other language can be held to the same contract without cloning anything. The data behind every page is in the section above as JSON, XML, SQL and a loaded SQLite database. The repository holds all of it.