Score a claim from your own assistant
The engine is exposed as a Model Context Protocol server at https://alltruefact.com/mcp (Streamable HTTP, no login). Any MCP client can list its tools and ask it to score a claim; the result is a page on this site plus the full JSON.
Connect
Claude Code:
claude mcp add --transport http alltruefact https://alltruefact.com/mcp
Claude Desktop, ChatGPT and other clients: add a remote MCP server (sometimes called a connector or custom integration) with the URL https://alltruefact.com/mcp. No authentication.
Tools
| Tool | What it does |
|---|---|
score_claim(text, strictness?) | Queues one declarative claim for scoring on the public profile with the locked defaults and returns a job id. Minutes, not seconds. A values claim is refused with the reason; identical claims are deduplicated. |
get_job(job_id) | Status of a job; when done, the scored result and the page URL. |
get_claim(slug, strictness?) | Any scored claim: consensus, per-model verdicts, tree, links to the page and the JSON. |
list_claims() | Everything scored on the site. |
The one dial
strictness, 0 to 1, default 0.5. It re-scores the same cached judgments with a harsher independence discount, a higher evidence floor and a steeper source-quality curve, and shows the result next to the verdict of record. It never replaces the verdict of record, so every claim on the site stays comparable. The mapping is on the rubric page.
Limits
- A scored claim costs the site about a dollar in model calls, so each address gets 3 new claims per UTC day and the tool as a whole stops accepting new claims once its daily spend cap is reached. Reading anything already scored is free and unlimited within the request rate.
- One claim scores at a time; a queue position is returned.
- Claims are one declarative sentence, 12 to 400 characters. Questions get rephrased by your assistant; give it the claim you want tested.
- Trees for submitted claims are written by the retrieval model and reviewed by nobody. Video claims have hand-written, reviewed trees. Both are shown in full on the page.
Right now: 0 queued, $0.00 committed today of a $6.00 cap. Scoring models: anthropic/claude-haiku-4.5, openai/gpt-5.4-mini, google/gemini-3.5-flash-lite, deepseek/deepseek-v4.1-flash, x-ai/grok-4.3; retrieval and roots by deepseek/deepseek-v4-pro-0813.
What you get back
The consensus label and probability, each model's label and probability with its cold opinion beside it, the decomposition tree with any excluded values parts and their reasons, and links to the page and the JSON. The JSON has every document, every judgment with its model string and prompt version, every cluster, and the parameter snapshot.