One deterministic path from agent to verified output.
Claim extraction and evidence mapping are the only stages that use an LLM. Everything after that, comparison, tolerance, scoring, is a pure function. The same document and evidence always produce the same Trust Report.
An agent produces or holds a document containing claims: a report, a summary, an API response.
Receives the document and its evidence sources through MCP or REST. Discovery is free; verification is paid.
The Excel or CSV files a claim should trace back to. An LLM maps claim terminology to evidence columns, nothing more.
Pure, deterministic comparison. Tolerance checking, severity classification, and trust scoring, with no LLM involved.
A Trust Report: a score, a per-claim breakdown, and an explanation for every mismatch.
The agent that receives the report and decides whether to act on the original claims.
Where the LLM boundary sits
Claim extraction and evidence mapping
Reading a document and finding the column in a spreadsheet that corresponds to "Revenue" requires semantic understanding. Both stages fail closed: an unparseable document produces zero claims, an unmatched claim is marked unmapped, never guessed.
Comparison, tolerance, and scoring
Once a claim is mapped to a value, everything from there is arithmetic: absolute and relative difference, severity classification against a metric-aware tolerance, and a weighted trust score. No model is involved, and no two runs can disagree.