checking-facts
Skill33 starsChanged 16 days agoMITView on GitHub
Checks factual assertions (numbers, quotes, dates, versions, API or tool behavior, what a source actually says) against primary sources read this pass, confirming each load-bearing fact by a second independent means. Use before a packet, doc, review, or release note relies on a fact. Do not use for citation-lineage wording (checking-source-claims) or mapping claims to local test evidence (proving-claims).
In this repository34 other files it gives its agents.
- AGENTS.mdAGENTS.md
- challenge/AGENTS.mdAGENTS.md
- challenge/CLAUDE.mdCLAUDE.md
- CLAUDE.mdCLAUDE.md
- breaking-down-the-workSkill
- briefing-an-agentSkill
- checking-legal-and-safety-wordingSkill
- checking-release-readinessSkill
- checking-source-claimsSkill
- checking-what-a-change-affectsSkill
- choosing-what-to-controlSkill
- closing-stale-packetsSkill
- creating-change-recordsSkill
- deciding-who-decidesSkill
- declaring-intentSkill
- double-checking-before-actingSkill
- handing-off-workSkill
- learning-from-experienceSkill
- organizing-project-foldersSkill
- proving-claimsSkill
- questioning-attitudeSkill
- rating-change-riskSkill
- recording-a-known-good-versionSkill
- recording-what-an-agent-didSkill
- reporting-shared-defectsSkill
- responding-to-incidentsSkill
- reviewing-code-qualitySkill
- running-a-review-panelSkill
- staying-on-missionSkill
- stress-testing-agent-changesSkill
- tracking-deficienciesSkill
- using-nuclear-gradeSkill
- verifying-final-artifactsSkill
- vetting-outside-code-and-modelsSkill
The file
We split this skill into the rules it gives an agent
5 rules. About 0 are common, 5 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈5
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 16
Claim checked: every load-bearing factual assertion traces to a primary source read this pass and is confirmed by a second independent …
- Distinctive: line 46
Read it this pass. Memory, training data, a snippet, an LLM summary (including your own earlier turn), and the citing doc's paraphrase are …
- Distinctive: line 60
Every load-bearing fact has a primary locator, a second means, and a verdict.
- Distinctive: line 61
No fact is cited to memory, a snippet, or a model summary.
- Distinctive: line 66
Stop when a load-bearing fact cannot be verified and the text still asserts it.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Checking Facts
Overview
A fluent sentence carrying a wrong number reads exactly like one carrying a right number. Agents state facts from memory, search snippets, a secondary summary, or their own earlier output, and all of those are leads, not sources. This skill traces each load-bearing fact to its primary source, reads that source now, and confirms the fact by a second means that cannot share the first's common cause, the alternate-method rung in verification-methods-and-depth.md. A fact that survives is stated with its scope and locator. One that does not is corrected, labeled, or removed.
Boundary: checking-source-claims asks whether a citation is worded honestly (influence, not compliance). proving-claims maps claims about this change to local evidence. This skill asks whether an assertion about the world is true.
Decision contract
- Claim checked: every load-bearing factual assertion traces to a primary source read this pass and is confirmed by a second independent means, and no unconfirmed fact is stated as settled.
- Artifact observed: the asserting text, each primary source at a locator, and the second means -> a fact ledger and corrected wording.
- Decision affected: block -- whether the text may rely on the fact as stated.
- Failure class: unverified-fact (asserted from memory, a summary, the actor's own output, or one source standing in for two).
- Next action: correct to what the source supports, label as unverified, or remove; a disputed fact under a trust-bearing decision escalates.
When to Use
- A packet, doc, review, release note, or answer states a number, quote, date, version, limit, price, or availability.
- Text describes how an API, library, CLI, model, or service behaves.
- Text says what a statute, standard, paper, issue, or vendor page says.
- A reviewer would act differently if the fact were wrong.
When Not to Use
- The question is whether citation wording overclaims; use
checking-source-claims. - The claim is about this change's behavior and a local test settles it; use
proving-claims. - The fact is not load-bearing and is already labeled as approximate or unverified.
Inputs
- The text and the decision that depends on it.
- Access to primary sources: official docs at the pinned version, source code, statute text, the paper itself, raw data, or the live system.
- The mode: Quick, Standard, or Nuclear.
Process
- Extract each assertion. Type it: quantity, quote, date or version, behavior, attribution, or current state.
- Mark the load-bearing ones: a decision, claim, or public statement fails if they are wrong. If there are more facts than you can check, define what counts as an error first, check every load-bearing fact, and sample the rest at random. Report a clean sample of n as an upper error bound of about 3/n, not as "all correct".
- Find the primary source, the origin rather than a page about it. Record a locator: URL or path, section or line, version, and date accessed.
- Read it this pass. Memory, training data, a snippet, an LLM summary (including your own earlier turn), and the citing doc's paraphrase are leads.
- Confirm by a second means that cannot share the first's common cause: an independent authoritative copy; running the code or calling the API against the docs; recomputing a number from raw data another way; an exact-string match for a quote against a second, independently published copy; a minimal reproduction at the pinned version. Re-reading the same copy is a check, not a second means. Two pages repeating one upstream are one source. Two model answers are one source.
- Record a verdict: confirmed, confirmed with qualification, contradicted, or unverifiable. Note the version or date the fact holds for.
- Rewrite to exactly what the sources support, with scope and locator. Present a conflict as a conflict. Label or remove what you could not verify.
- Grade by who does the second means. Quick: the author may do both. Standard: someone other than the author does or re-runs the second means. Nuclear: that person is also briefed without the author's answer.
Outputs
- A fact ledger in
verification.mdor beside the doc: fact | type | primary source and locator | second means | verdict | as-of. - Corrected wording.
- Unverified or disputed facts carried as named gaps.
Verification
- Every load-bearing fact has a primary locator, a second means, and a verdict.
- No fact is cited to memory, a snippet, or a model summary.
- Quoted text matches the source exactly; volatile facts carry an as-of date.
Escalation
- Stop when a load-bearing fact cannot be verified and the text still asserts it.
- Escalate when sources conflict and the decision depends on which is right, or when the fact bears on public trust, safety, security, legal, or procurement.
- Return unverifiable, not a guess, when the primary source is unreachable here. Name the gap.
Common Rationalizations
- "I know this." Memory is a lead; read the source.
- "Two sources agree." Check they do not share one upstream.
- "A second model confirmed it." LLM errors are strongly correlated across families; that is partial diversity at best.
- "Every fact I checked was right." Four clean checks bound the error rate only below about 53%.
- "It is in our own docs." Your docs can carry a propagated error.
- "The docs say so." Docs drift from behavior; for behavior, run it.
- "It is a minor number." Then label it; do not assert it.
- "It is just the citation." Authors, dates, and titles are facts too; read them from the record.
Red Flags
- A fact with no locator, or a locator to a summary rather than the origin.
- Paraphrase inside quotation marks.
- Version-dependent behavior with no version; a price, limit, or availability with no as-of date.
- The only confirmation is the actor's own earlier output.
Prompt
Run a Nuclear-grade fact check.
Inputs:
- text and the decision that relies on it: <paste/link>
- mode: Quick / Standard / Nuclear
- primary sources available: <docs at version / code / statute / paper / data / live system>
Do this:
- Extract each assertion; type it; mark the load-bearing ones.
- Trace each to its primary source and read it this pass; record a locator.
- Confirm each load-bearing fact by a second means that cannot share the first's common cause.
- Rewrite to what the sources support, with scope and as-of date.
Return:
- a fact ledger: fact | type | primary source and locator | second means | verdict | as-of
- corrected wording
- unverified or disputed facts as named gaps, and any escalation
Do not treat memory, a snippet, or a model summary as a source. Do not imply formal verification, compliance, certification, safety, security, or regulatory adequacy.Source-lineage note
An original fact-checking workflow, not a claim to have invented source verification. Concept lineage: alternate or simplified calculational methods as one way of verifying or checking a design, by someone other than the originator (10 CFR 50 Appendix B, Criterion III), and independent verification as "freedom of thought from the performer" plus validating assumptions (DOE-HDBK-1028-2009), both mapped in docs/00-standards-foundation/source-map.md. It does not create formal verification and validation, compliance, certification, or any safety, security, or regulatory guarantee.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

