Home > Page Design > ReasonRank

ReasonRank: The PageRank for Arguments

ReasonRank is a network-based scoring algorithm that measures the strength of a belief by the quality of its supporting arguments, not its popularity. Google's PageRank treated hyperlinks as votes for a page's importance. ReasonRank treats logical arguments as votes for a belief's validity, with one crucial difference: a vote here is only worth what its own evidence and reasoning can defend.

The Problem vs. The Solution

How ReasonRank Compares to PageRank

ReasonRank applies graph theory to logic. Every claim is a node, every reason is a weighted link, and strength flows through the network.

The Scoring Dimensions

Every argument's contribution to a belief runs through four factors, and a belief's score is the sum of supporting contributions minus opposing ones, computed recursively, because every argument is itself a belief with its own tree of pros and cons underneath it. Here is the conceptual shape:

An argument's contribution ≈ Evidence Quality × Linkage × Importance × Uniqueness

That is the readable summary, not the exact math. The exact math, the full recursion, normalization, and every factor's equation, is REASONRANK_EQUATIONS.md in the repo, with each equation tied to the file that implements it and the tests that pin it down.

1. Evidence Quality (EVS)

How good is the backing? The Evidence Verification Score grades each row on its tier, its replication count and consistency, and its relevance. The live rule for an evidence row's impact is sign × EVS × |linkage| × 100, computed by the engine on every propagation pass, never typed in by an editor. On the seeded UBI page, verified and replicated studies carry about 78 points each while an unverified survey carries 3. That is quality ordering by structure, not by whoever edited last. Reference: Evidence

2. Linkage

If this were true, would it actually move the conclusion? Linkage is the relevance gate, and it is what makes true-but-irrelevant claims contribute nothing no matter how loudly they are repeated. Reference: Linkage Scores

3. Importance

Does this point address a core pillar of the question or a triviality? Importance weights each consideration by how much rides on it, anchored by the objective criteria debate for the belief. Reference: Importance Score

4. Uniqueness

Is this a new point or the same point in a new costume? Semantically similar arguments are grouped, and a restatement contributes only its novel fraction, so nobody inflates a score by saying one thing five ways. The discount runs in the engine's impact calculation. Reference: Belief Sorting and Uniqueness

Logical validity is scored on its own track rather than blended into evidence, and a fallacy accusation is not an automatic penalty: it is itself an argument, scored like any other, attached to the specific factor it damages. A relevance fallacy attacks Linkage, a formal fallacy attacks Logical Validity, cherry-picking attacks Evidence Quality. Reference: the bias and fallacy catalog

The Circuit Breaker: Retraction That Propagates

Debunk an argument's data and every conclusion built on it collapses automatically. Here is the machinery.

Every evidence row carries a verification standing: VERIFIED, UNVERIFIED, DISPUTED, or FALSIFIED. A verified row carries full weight. A disputed row is cut to half until the dispute settles. A falsified row drops to zero, a circuit breaker for the whole chain above it. Changing a row's standing goes through a single API call that requires a rationale, logs it to the audit trail, and propagates the change through every score that leaned on that row. The Evidence Ledger on each belief page shows the standing inline, with falsified rows struck through.

In the seeded corpus, marking the Finland pilot study falsified takes its impact from 78.4 to 0 and drops the UBI belief's net from +47.5 to +33.8 in one call, rationale on the audit log; restore the row and the score comes back. No editor went around updating talking points. The graph did it, because in a network of logic, a dead foundation cannot keep holding up a roof.

What Kind of Fight Is This?

Scoring the arguments buys you something subtler than a winner: a diagnosis. The conflict resolution pipeline reads the scored rows and classifies what kind of disagreement it is looking at. If the evidence weight is genuinely two-sided, it is a factual dispute, and more data can still move it. If the action lives in the relevance links, it is a linkage dispute, and it resolves in the sub-debates about whether the facts actually bear on the conclusion. If the widest gap is in how the two sides rank their values, it is a values conflict, and it needs compromise design, not another study. Close races honestly read as mixed.

In the seeded corpus, the UBI page diagnoses as a values conflict, which anyone who has argued about UBI at a dinner table will find plausible. The classifier and the shared-interest detector are deliberately simple, legible heuristics over the scored rows; they sharpen as the graph fills.

The System Scores Its Own Thesis

The claim this page makes, that arguments should be ranked by logical support, is itself a belief in the corpus, with its seven supporting reasons seeded as scored child beliefs. And because the method demands it, the con column is not empty. The two strongest objections we know of are on the page and scored: metric gaming (Goodhart's law, make a measure a target and the measure decays) and garbage-in (a perfect calculator over a junk graph returns junk). The evidence rows carry real provenance, Brin and Page's 1998 PageRank paper on one side, the Goodhart literature on the other.

In the seeded corpus, the agree case holds 81 percent of the argument weight, a 61-point margin. A young corpus, not a coronation. If you think the thesis is wrong, the highest-leverage rebuttal is not a comment anywhere on the internet. It is a strong argument in that con column, because the score has no choice but to follow it.

Also Wired In

• Linkage propagation: scores flow through the entire graph on every pass, so a change anywhere updates everything downstream.
• Objective criteria: the debate over how to measure a belief, feeding the Importance weights.
• The denominator: beliefs are scored against their rivals, not in a vacuum. Canonical spec is THE_DENOMINATOR.md in the repo.

Ready for Better Thinking?

The map is REASONRANK_EQUATIONS.md (every equation, its implementation file, and its tests), and there are more good first issues than hands. Clone it, run it, pick a function. Or skip the code and add a strong argument to a belief page, because a thin corpus is the one thing the algorithm cannot fix for itself.
View the Code on GitHub | Contact the Project Lead