Evidence & Citation Checklist

If a claim in your report can't be traced to a source, it isn't a finding — it's a hope. This checklist turns that standard into fourteen binary checks you can run on any draft in about five minutes, and it hands you the result as a report you can paste into a revision request.

Who this is for

Use it when a report was drafted fast — by you, by a colleague, or by a model — and someone downstream will act on it: a budget line, a hire, a shipment. It fits executive reports, decision memos, and research summaries.

It is not for academic citation formats (this checks whether evidence exists, not whether your bibliography is styled correctly), and it is not a plagiarism scanner — it never compares your text against the web.

What the fourteen checks cover

The items are the evidence bar we apply to our own reports, generalized:

• Sources exist (3). Every factual claim maps to a named source, every link opens, every source is labeled by what it is. A claim with no source is the single most common defect in AI-drafted reports.
• Quotes and numbers are real (4). Quotes checked verbatim, in context; numbers traced; dates spot-checked; confidence labels that say why. The verbatim check exists because we watched it fail: in our own field run, the model asserted an input excerpt was truncated when the source file was complete — a confident, plausible, false citation claim.
• Gaps are declared (3). A "what we did NOT verify" list; conflicts surfaced instead of averaged away; inferences labeled as inferences.
• Provenance is traceable (3). AI passages read in full by a human; production method and revision recorded; inputs linked.
• It survives leaving your tool (1). The exported file opens cleanly somewhere else — the portability smoke test.

Input it expects

You answer for the draft as it stands — fourteen yes/no attestation checkboxes. The tool reads nothing and stores nothing; the result exists only in your browser until you copy or carry it away.

Example result

Running the audit against a draft that fails three checks produces this report (plain text, ready to paste):

11 of 14 checks passed. Fix before shipping:
Failed items:
- Direct quotes were checked verbatim against the source text, in context. → Re-read every quote against its source; fix drift.
- A "what we did NOT verify" list exists — a bare "all verified" fails. → List what was not checked and why it still holds.
- The exported file opens cleanly outside the tool that made it. → Open the export once in a second tool; note what breaks.
Audited with the AtomStorm Evidence & Citation Checklist (atomstorm.ai/en/tools/evidence-citation-checklist/)

Each failed line pairs the check with its fix. "Copy report" puts that text on your clipboard; the button next to it carries the same text into AtomStorm Studio as the brief for a revision pass, so the model fixes the named failures instead of regenerating blindly.

What we verified, and what we did not

• Verified: the fourteen checks are the same bar our editorial checklist (18-item acceptance checklist) applies to our published reports, plus three failure modes our independent review actually caught — verbatim-quote drift, bare confidence words, and contradictions averaged away. The tool logic is unit-tested; the result text is capped at the same 2,000-character limit the Studio handoff channel enforces.
• Not verified: we have not run usability sessions with external teams, and the audit is only as honest as your answers — attestation, not analysis. If your draft would fail an item, checking it anyway produces a clean report for a dirty draft.