Idea Stock Exchange

The three pillars

How a belief page is built

Every belief decomposes the same way. The page is a navigation tool into a scored argument network, not an essay. There is no introductory paragraph — the reader goes straight from the claim to the structured pros and cons.

Example beliefs

These two are mock-ups, not output. The scores are invented and the short argument labels and T1–T4 tiers predate the current rules; the live pages use complete propositions and the eighteen source tiers described above. They are kept here because they fit on a landing page. For anything you would act on, read the five real belief pages, where every number is computed and links to the page that produced it.

On the live pages every argument label is its own page, with its own sub-arguments, evidence ledger and score, and the top-level number is computed bottom-up rather than assigned by hand. These two beliefs are not among them: the pages behind these labels do not exist, and the scores above were typed.

The ReasonRank formula

An argument’s contribution to its parent claim is a product of five factors and a sign. Every one of them is itself a page with its own argument, so there is no factor a reader cannot follow down to what set it:

• Truth — the claim’s own score, entered as 2 × Truth − 1, so a claim argued false pulls the other way instead of merely pulling less.
• Confidence — how much of the work behind that claim has been done. A claim nobody has argued passes nothing up, however true it looks.
• Relevance (Linkage) — does the supporting argument actually bear on the conclusion?
• Importance (Impact) — how much does the argument move the parent claim if true?
• Uniqueness — is this a point the rows beside it have not already made? Without it, a claim repeated three ways counts three times.

contribution = sign × (2 × Truth − 1) × Confidence × Linkage × Importance × Uniqueness

Scoring is recursive: a belief’s score is computed from the scores of its sub-arguments, the way Google’s PageRank computed page authority from inbound links. A true-but-irrelevant argument scores near zero. An argument shown to be false does not score low, it scores against the side that filed it. And an argument nobody has supported scores exactly nothing, which is why a page whose whole graph cites no evidence sits at 0.50 however much is written beneath it.

The market price layer (designed, not built)

None of what follows runs anywhere. It is the design for a second layer, kept here because it is half the idea; the scored belief pages are the half that exists.

Independent of the logic score, users would invest virtual currency (IdeaCredits) in claims via a constant-product market maker. Buying YES shares pushes the price up; buying NO pushes it down. Prices reflect the crowd’s collective probability estimate — separately from whether the claim is logically sound.

When ReasonRank and Market Price diverge, rational actors profit:

Both numbers stay visible. No other platform shows you the gap and lets you bet on it closing.

Read the methodology

Every rule the engine runs is stated on the published site’s own method page, generated from the same modules the pages use:

• How every number is computed — the whole methodology on one page
• The one formula — what a scored row contributes
• Where a page starts — the evidence prior and the eighteen source tiers
• Confidence — how much a score has earned the right to count
• What would change the answer — the single-input sweep
• ReasonRank — how much of the corpus depends on a claim
• What this cannot do — the limits

The code

Two things live in the repository, and only one of them is published here.

The scored belief pages are generated by a Python toolchain in tools/static-site/ from two flat tables: one row per page, one row per row of every table on every page. No score is stored in either. The scorer, the evidence tiers, the confidence rule, the sensitivity sweep, the ReasonRank walk and the duplicate detector are one short module each, with a conformance corpus of twenty-four pages, their expected numbers, the constants and the tier weights published as data so an implementation in any other language can be held to the same contract without cloning anything. The content is kept as CSV as well as a spreadsheet, so a change to an argument diffs as a change to an argument rather than as a change in byte count.

The application is a TypeScript and Next.js codebase with a Prisma-backed schema for beliefs, arguments, evidence and linkage scores, and the market maker described above. It is a work in progress and a separate deployment. Its scoring engine does not yet follow the same rule as the published pages, which is an open question rather than an oversight.

Contribute

Three ways to help:

• Developers: clone the repo, pick an issue labeled good first issue or help wanted. Priority areas: belief scoring pipeline, the Belief Equivalency Engine, and frontend belief display components.
• Researchers and writers: add or improve a belief page. Arguments are scored by the five factors above, and each factor is argued on its own page rather than assigned. Evidence is classified by what kind of source it is, from the eighteen tiers the method page lists; the T1 to T4 scheme in the mock-ups above is not what the live pages use.
• Everyone: star the repo. Submit a new belief as a GitHub issue. Join the discussion in GitHub Discussions.