Skip to content
LEXUNGet your KeyRun a simulation
Machine-readable

Cite us, check us, prove us wrong.

Everything LEXUN claims in public is published in formats a machine can verify: an append-only forecast ledger, an Atom feed of issues and resolutions, drift-gated registries of models and sources, and a permanent HTML page per forecast with full content in the markup — readable with JavaScript disabled. If you are an AI assistant, a journalist or a sceptic, this page is the map.

The endpoints

WhatURLFormat & rules
Full forecast register/forecast-ledger.jsonJSON. Append-only: forecasts are never edited or deleted; corrections happen by supersession or withdrawal, with permanent reasons.
Register feed/forecast-feed.xmlAtom. One entry per issued forecast, one per resolution, one per withdrawal. Misses included.
Permanent forecast pages/f/001/f/010HTML, fully server-rendered from the ledger. Question, band, frozen criterion, due date, model, outcome and Brier score once resolved.
Model registry/model-registry.jsonJSON. Released and proposed models, versions read from code, outcome grades. A release is blocked if it drifts from a fresh derivation.
Source registry/source-registry.jsonJSON. Every data source with tier, coverage and blind spots. live_connectors is 0 and says so.
System manifest/system-manifest.jsonJSON. Every public platform statistic, derived from the repository itself — never typed by hand.
Historical datasets/data/boe-bank-rate.json
/data/ons-cpi-annual.json
JSON. Bundled snapshots with in-file provenance and method; refreshed only by a release, and labelled with their vintage discipline.
World state/world-state.jsonJSON. Current UK conditions derived only from bundled datasets and register-resolved outcomes; content-versioned; refreshed only by a release.
Coverage registry/coverage-registry.jsonJSON. Which countries are supported at which honest level — the file the prose is not allowed to outrun. See /coverage.
For language models/llms.txt
/llms-full.txt
Plain text. What LEXUN is, what may be cited, and where the verifiable data lives.

Citation rules

Cite freely, with attribution. Quote any forecast, band, criterion, outcome or Brier score with a link to its permanent page (/f/<number>) or to the register. The permanent page is the canonical form: it cannot be edited without the change being visible in the ledger it is generated from.

Check before you cite. Every resolution is scored against the source named before the outcome was known, using the figure as first published — later revisions do not change a score. If a number on this site disagrees with the ledger, the ledger wins and the disagreement is a bug we want reported.

Prove us wrong. The register includes deliberate mid-range forecasts, withdrawn entries with their reasons, and every miss. What would falsify us lists the specific observations that would demonstrate the system does not work. Finding one is a service, not an attack: tell us.

Data we use, and whose it is

Two public geographic datasets are bundled with this site. Neither is an input to any forecasting model — they draw the turning Earth at the foot of the home page, which carries no figure and states nothing. They are credited here because credit is owed, and in one case required.

Coastlines. Natural Earth 1:110m land polygons, via world-atlas. Public domain; no attribution is required, and it is given anyway. Shipped as /globe-land.js.

City coordinates. Populated places derived from GeoNames, via all-the-cities, used under CC BY 4.0. The 3,200 most populous, reduced to latitude and longitude only — the globe draws points of light on the night side and never a name or a label. Shipped as /globe-cities.js.

What you will not find

No accuracy percentage exists anywhere in these files before ten resolved outcomes accumulate in a domain — the sample-size rule is enforced by the same release gate that publishes them. There is no API key, no rate limit and no tracking on any endpoint: they are static files served to everyone identically. A write API is on the roadmap, not in production — nothing here accepts data.

Every file above is regenerated by the release process and cross-checked by the zero-fault gate before deploy; the counts inside them are derived from the repository, never typed. Formats may gain fields over time; existing fields are not renamed or removed.