Evidence Cards for Research: A Claim-to-Source System

Folders organize files; evidence cards organize what those files can support. Write one precise proposition at the top of each card. “The framework has four functions” is reviewable. “Useful points from the report” is not. If a sentence contains two claims that could have different evidence, split it.

The card should state whose claim it is. A source claim reports what the author says. A researcher's inference connects evidence to a conclusion. A hypothesis identifies something to test. Keeping those types explicit prevents generated interpretation from being mistaken for language in the source.

Store a stable card ID, claim, claim type, source ID, source title, locator, short support, scope or conditions, relationship, verification status, reviewer, and checked date. Add conflict IDs and the draft sections that use the card. The schema looks detailed, but most fields are short and remove detective work later.

NIST's work on assurance cases describes evidence as support for claims and emphasizes traceability to the source, along with assumptions and conditions. Content research is not a software assurance case, but the same logic is useful: a persuasive claim needs a visible argument and evidence trail, not merely a confident paragraph.

During reading, create candidate cards quickly and preserve locators. AI can propose claim wording, identify a nearby qualification, or format the fields. Keep those cards in a candidate lane until the original source is reopened. The model's summary belongs in a working field, not in the verified-support field.

Avoid copying long passages. Capture the minimum text needed to check the interpretation, and return to the source for wider context. If exact wording matters, use a short quotation with its locator; otherwise write a faithful paraphrase and keep it visibly marked as yours.

A passage can be quoted accurately and still fail to support the claim. Check subject, time, population, modality, method, and nearby exceptions. Ask whether the card narrows or strengthens the source. Then select supports, qualifies, conflicts, exemplifies, or does not address, and write one sentence explaining the choice.

Mark uncertainty instead of resolving it through phrasing. If OCR is doubtful, if a transcript speaker is unclear, or if a table cell cannot be read, the card remains needs review. The downstream draft can see the candidate without treating it as accepted evidence.

When two cards disagree, preserve both and create a conflict record. State the exact proposition, each source's scope, the likely cause of difference, and what would resolve it. Two sources published in different years may both be accurate for their periods; two studies may use different definitions; one source may simply be wrong.

Do not average incompatible claims or ask AI for a compromise sentence. The draft can present both positions, choose one with an explicit rationale, or narrow the claim to the overlap. The conflict record keeps that editorial decision reviewable.

Use candidate, checked, accepted, rejected, stale, and needs review. Accepted means the source and relationship were checked, not that the claim is eternally true. Add update triggers for current facts such as product behavior, policies, prices, and model labels. When a source changes, mark dependent cards stale before revising the draft.

Before publication, follow every material sentence to its cards and every accepted card to its source. After publication, keep the card-to-draft links so an update can find affected passages. A reusable evidence system earns its value during revision, when memory is weakest and the original research session is gone.