synthesis-fact-checking

Fact-Check Process for Articles and AI-Synthesized Content

A repeatable process for verifying the factual accuracy of articles, blog posts, news and AI-synthesized content before publication. It is calibrated for recursive contamination: AI output now feeds the sources the next AI and the next researcher cite, so agreement among sources counts only when they are independent. synthesis-content-quality covers style and substance; this skill covers factual correctness.

Binding rules

Rule numbers are the section numbers that other files cite as "SKILL.md section N" (section 2, 4a to 4g, sections 7 and 8). Contents names the file that holds each section in full.

1. Extract every verifiable claim before verifying any: numbers, studies, authors, dates, quotes, organizations, causal and comparative claims, URLs, DOIs, nested attributions. Claims that "sound right" are often the subtly wrong ones.
2. Agreement counts only among graph-independent sources. Sources that trace to one AI-generated upstream, or to the author's own doctrine, are single-sourced however many there are.
3. Verify against the primary source. Secondary and tertiary sources help you find it; AI output is what you check, never what you check against.
4. Check the framing, not just the number. The common errors are 4a to 4g: wrong framing, conflated findings, wrong specifics, wrong entity, misattribution, fabricated citation, superseded data.
5. Apply the nine C1 protocols wherever their claim type appears: nested attribution, paraphrase drift, composite quotes, position shifting, translation drift, URL rot, synthetic sources, citation laundering, per-family hallucination.
6. When the producing model family is known, run its signature check first (Claude: DOIs; GPT: URLs; Gemini: vague attribution), rather than every check on every piece.
7. A quote is verified only when it exists, its wording is exact, its speaker and context are right, and it is one continuous utterance traced to the original speaker; paraphrase never gains quotation marks. A misquote damages trust out of proportion.
8. A study is verified only when existence, paper number, every author, venue, method, findings and retraction status check out. An unfindable study's citation is removed; its claim is re-sourced, hedged or cut.
9. A backdated article cites only sources that predate its date, unless marked "Updated on", and its time-relative words are true as of that date.
10. After any de-jargoning, anonymization or accessibility pass, re-verify every replaced term. Accessible wording can stop being true.
11. Log every claim checked in a review-log.md beside the draft, with status, source, finding and correction.
12. Nothing publishes until the pre-publish checklist passes.
13. Route each claim through the quick decision tree. Opinion needs no check; a claim you cannot source is re-sourced or hedged if essential, removed if not.

Contents

• references/verification-steps.md: sections 1, 2, 3, 7, 8 and 13 in full: claim extraction, the confidence table and graph-independence check, the source hierarchy, the quote and study protocols, the decision tree. Read it when starting a fact-check.
• references/error-patterns.md: sections 4 to 6: errors 4a to 4g with examples, summaries of the nine C1 protocols and of the family signatures. Read it when a claim looks wrong, to name the error and pick the deeper protocol.
• references/temporal-and-translation.md: sections 9 and 10. Read it when the article is backdated or has been de-jargoned, anonymized or simplified.
• references/review-log-and-checklist.md: sections 11 and 12, the review-log template and pre-publish checklist. Read it when logging claims and before publishing.
• detailed-protocols.md: All nine C1 protocol sections with full failure modes, worked examples, and detection procedures. An index to three parts; read the part with the protocol you are applying:
  • references/detailed-protocols-quotes.md: C1-NESTED, PARAPH, COMPOSITE, POSSHIFT
  • references/detailed-protocols-sources.md: C1-TRANS, URLROT, SYNTH, LAUNDER
  • references/detailed-protocols-model-families.md: C1-TOOLHALL, incident list
• autopilot-research-quality.md: Decision-changing primary evidence, counterevidence and task-specific research acceptance. Read it when research must support a decision.
• per-family-hallucination-signatures.md: Detailed per-family signature catalog with empirical anchors. An index to two parts; read it for rule 6 in depth:
  • references/per-family-signatures-claude-gpt-gemini.md: why, caveats, anchors; Claude, GPT, Gemini
  • references/per-family-signatures-llama-grok-deepseek-mistral-qwen.md: Llama, Grok, DeepSeek, Mistral, Qwen
• citation-laundering-detection.md: Graph-traversal protocol for detecting citation laundering chains. Read it when sources may share an upstream.
• production-incident-archive.md: Documented 2024-2026 production incidents (Mostafavi, Goldberg Segalla, Chicago Sun-Times, Springer book, BBC/EBU, Topaz Lancet, and others) with detection lessons. Read it for precedent.
• bibliography.md: Consolidated bibliography with verification status. Read it before repeating a figure from this skill.
• references/background.md: purpose, what v2.0 added, related skills. Read it once, or when choosing between this skill and a sibling.
• references/coverage-map.md: where each part of the 2.1.1 text now lives.
• Process Overview: below.

Process Overview

1. Claim Extraction
2. Multi-Source Confidence Assessment (with graph independence per v2.0)
3. Verification by Source Hierarchy
4. Common Error Pattern Detection (4a through 4g, refreshed)
5. Nine new protocol sections for v2.0 (C1)
6. Per-family hallucination signatures (v2.0)
7. Quote Verification (with nested attribution per v2.0)
8. Study Verification
9. Temporal Verification (if backdated)
10. Translation-Pass Re-Verification
11. Documentation
12. Pre-Publish Checklist