Rseng Fact Checking

Fact-checking and source verification

A reference can exist, resolve and be unretracted - and still not say what it is cited for. That last mile is where fabrication actually lives: real papers attached to claims they do not support, secondary sources laundered as primary, numbers that drifted in retelling. rseng-citation-hygiene verifies that references EXIST and are current; this skill verifies that they SUPPORT their claims and deserve trust. The duty is self-referential and stated plainly: an agent that asserts a fact or attaches a source must be prepared to show where in the source the support lives - and must say "I could not verify this" when it cannot. Confidence is not evidence.

Claim-source alignment: the core check

For each claim-citation pair:

1. Extract what the claim actually asserts: the specific finding,

number, method or position attributed to the source - not the general topic. Most misattribution hides in specificity ("X causes Y" cited to a paper showing correlation; "widely used" cited to a paper that merely mentions).

1. Open the source and FIND the support: the abstract often

suffices for existence of a finding (OpenAlex serves abstracts); exact numbers, conditions and caveats need the relevant section. Record WHERE the support is (section, table)

• unfindable support is the finding.

1. Grade the alignment honestly: supports as stated / supports

with caveats the text drops / related but does not support / contradicts / cannot access to verify. The middle grades matter most - dropped caveats ("in mice", "n=12", "simulated data") are the commonest integrity leak (rseng-research-integrity's spirit applied to prose).

1. Check the chain: if the source itself cites something for the

claim, it is a secondary source - prefer citing the primary after checking IT (citation chains degrade like photocopies; "as cited in" is honest when the primary is unreachable).

Batch mode for documents: extract all claim-citation pairs, run the alignment check per pair, and report a table with grades and locations - the content-level companion to rseng-citation-hygiene's existence table, and a natural pre-submission step (rseng-research-integrity's battery).

Source trust assessment

Not all resolvable sources deserve citation:

• Status: peer-reviewed publication, preprint (legitimate but

label it - rseng-open-science-practices), dataset, documentation, blog post? Cite the thing appropriate to the claim: software behavior -> the docs; a finding -> the paper; never a press release for a scientific result.

• Venue signals: is the journal in DOAJ (for open access

venues), does the publisher follow recognizable review practice (Think-Check-Submit's checklist operationalizes this)? Predatory-venue signatures: guaranteed fast review, fake metrics, scope covering everything. A paper is not wrong for appearing in a weak venue - but weight it accordingly and prefer stronger support when available.

• Primary vs secondary: for facts, prefer the origin - the

standard's text (as this pack's own skills link eur-lex and specifications), the tool's documentation, the original study. Reviews are excellent context and honest citations for consensus claims - "reviewed in [X]" - not for specific findings.

• Currency: is the source still the state of knowledge? A

superseded version of a spec or a pre-replication-crisis finding may need updating, not just citing (rseng-citation-hygiene's preprint-to-published check is the mechanical half).

Verifying facts the agent asserts

The proactive duty, in practice:

• Checkable claims in generated text (numbers, dates, "X supports

Y", API behavior, legal obligations) get verified against an authoritative source before delivery, or labeled as unverified. For software claims, the executable check beats the document: run the command, read the actual API response (rseng-testing's instinct applied to prose).

• Distinguish knowledge from verification in the output when it

matters: "per the docs at " vs "from memory - verify before relying on this". Users calibrate on this honesty.

• When verification is impossible (paywall, offline, no

authoritative source), say so and mark the claim - an honest gap outperforms confident invention every time.

• Everything this pack already mandates applies: verified links

only (each skill's Learn-more discipline), retraction screening (rseng-citation-hygiene), and AI-contribution disclosure (rseng-ai-declaration) for documents the agent helped write.

Working with this skill

This skill is source-independent: it encodes claim-source verification practice, with the linked services as the checking infrastructure. It completes the verification chain: rseng-citation-hygiene (the reference is real) -> this skill (the reference supports the claim, and deserves to) -> rseng-research-integrity (the document's own numbers are consistent).

Learn more (verified):

• https://help.openalex.org - OpenAlex (metadata and abstracts

for alignment checks)

• https://api.crossref.org/swagger-ui/index.html - Crossref

REST API

• https://doaj.org - Directory of Open Access Journals
• https://thinkchecksubmit.org - Think. Check. Submit. venue

checklist

• https://gitlab.com/crossref/retraction-watch-data -

Retraction Watch database

Related skills

Check whether any of these applies before moving on:

• rseng-citation-hygiene - existence check runs first
• rseng-documentation - README claims need the same bar
• rseng-honesty - when asked to fake support
• rseng-open-science-practices - preprint status labeling
• rseng-research-integrity - document-level pre-submission battery
• rseng-testing - executable checks beat document claims

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: fdiblen
• Source: fdiblen/rseng-agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.