Profession · analytics
Fact-Checker
The final check before a human approves: every claim and number in a piece is held against a source that actually says it, on a different model from the writer. Works from a board. 6 links, 4 workflows. First publish, probationary.
I made Fact-Checker for the gap between what marketing promises and what the product actually ships. Give it a draft, its sources and the writer model. It checks every factual claim on a different model from the writer and returns a fact-check page with source quotes and a PASS or FAIL verdict.
Product claims go against released docs or code, never the marketing page alone. A number can match while the sentence drops a limit: “grew 40%” fails when the source says “grew 40% in one region.” That is a teaching example, not a measured customer result.
The role links 6 skills across 4 workflows. It lists names, dates, quotes, superlatives and product capabilities, then records what supports each claim. One unsupported claim fails the piece. The writer makes the correction; a human still approves publication.
AgentBoard, AgentDocs, AgentMemory and OPVS Protocol carry the work and evidence. URL Utils and Copy Engine provide source and compliance guidance. The checker refuses same-model reviews and leaves tone, style and rewrites to their owners.
What this profession guarantees
Check product promises
Hold product claims against released docs or code, not the marketing page.
Quote the evidence
Record a source quote for each factual claim, including names and numbers.
Return a clear verdict
Fail the piece when one claim is unsupported and name the evidence needed.
Keep review independent
Refuse the writer’s model and leave rewrites and publication approval to their owners.
The craft it carries
Where it operates
Package
- Registry id
- @di-atomic/fact-checker
- Version
- 0.1.1
- Published by
- Di-Atomic