Skip to content

occam_claim_check

Status: verified against src/FFOccamMcp.Core/Tools/OccamClaimCheckTool.cs @ 2026-09-29
Profiles: researcher, auditor, full — not default reader, minimal, or basic
Contract depth: MCP API · runtime tools/list

Purpose

Live claim grounding for one sentence against one URL: fetch via transcode pipeline, rank blocks by BM25 vs claim, return top matches with Merkle leaf + proof. Shows which source text was retrieved; does not decide support vs refute (verdict always not_evaluated).

When to use / when not

Use when Do not use when
Provable block(s) from this URL for one assertion Host-side stance status → occam_attest
You will judge support yourself from matches[].text Full page markdown only → occam_transcode
Re-check leaf + proof via occam_verify citation Expecting truth from the claim string alone

Parameters

Name Required Default Meaning
claim yes — Assertion to ground (a sentence)
url yes — HTTP(S) URL
backend_policy no http_then_browser http, browser, or http_then_browser
session_profile no null Optional session profile id
max_matches no 3 Max blocks (1–10; silently clamped)

Fixed internally: json_blocks=true, playbook_policy=auto; no caller max_tokens / fit_markdown.

Response shape

Success: found, retrieved (= found), verdict: not_evaluated, optional proven when found:false (retrieval-complete negative only — not semantic absence), blockMerkleRoot, matches[] (blockIndex, text, score, leaf, proof[], …), optional receipt, timestamp.

Failure: failure: { code, message } — page content unknown. Distinct from success with found: false.

Common failure codes

invalid_arguments; plus transcode taxonomy on fetch/extract. Full list: failure-codes.

OCCAM_PROFILE, OCCAM_RECEIPTS, session profiles — configuration.

See also