Claims scored from evidence, with the working shown.
A claim is split into checkable parts. For each part the engine searches for documents on both sides, asks five language models a series of narrow questions about every document, groups documents that rest on the same source so a pile of articles counts once, and adds the pushes in log-odds. The consensus, the divergence between the models, the model's cold opinion with no evidence, and the full scored JSON are published together.
The document judgments are still a model's opinion, made one narrow call at a time and recorded. Read how it works and what is still judgment.
Published claims
| Claim | Consensus | p | Models | Divergence | Scored |
|---|---|---|---|---|---|
| The Federal Open Market Committee raised the target range for the federal funds rate at its meeting of 15-16 September 2026. Post-cutoff backtest claim, five models, second pass fired. |
Supported | 0.999 | 5 | 1.83 | 2026-09-17 |
| Viktor Orban remained Prime Minister of Hungary after the Hungarian parliamentary election of 12 April 2026. Post-cutoff backtest claim, five models. |
Refuted | 0.000 | 5 | 6.97 | 2026-09-17 |
| The 2007 increase in US troop levels in Iraq, the surge, reduced violence in Iraq. Mirrored pair 3, claim A (Phase 3 human gate). |
Supported | 0.851 | 5 | 1.41 | 2026-09-17 |
| The 2009 American Recovery and Reinvestment Act, the stimulus, reduced US unemployment below what it would otherwise have been. Mirrored pair 3, claim B (Phase 3 human gate). |
Supported | 0.852 | 5 | 1.54 | 2026-09-17 |
| Iraq possessed weapons of mass destruction in March 2003, and that justified the invasion. The fixture claim; pre-cutoff for every model, so the cold numbers are recall. |
Refuted | 0.027 | 5 | 2.07 | 2026-09-17 |
| Immigration is good. A values claim. Refused with zero calls; the refusal is the published result. |
Refused (values claim) | – | 1 | – | 2026-09-16 |
Divergence is the mean pairwise difference in log-odds between the models' verdicts on the same evidence. Larger means the models read the same documents differently.
Other scored claims
Backtest runs and claims submitted through the tool. Same engine, same defaults; not on the curated list above. Trees for submitted claims are written by the retrieval model and reviewed by nobody.
| Claim | Consensus | p | Models | Scored |
|---|---|---|---|---|
| Immigration is good for a country. | Refused (values claim) | – | 5 | 2026-09-19 |
Score your own claim
The engine is exposed as an MCP server at https://alltruefact.com/mcp. Connect it to Claude, ChatGPT or any MCP client and ask it to score a claim; the tool queues the run, returns a job id, and the finished page appears here. Setup and limits.