How to Separate a Claim, the Evidence, and the Inference Between Them

A claim is not evidence. Evidence is not the conclusion. The inference is the bridge between them, and the bridge is where assumptions hide. If those three parts stay fused, confident language can make weak reasoning feel complete. Separate them and the real question becomes visible: does this evidence, under these conditions, support this claim more strongly than the alternatives?

This protocol works for research papers, product pages, AI answers, news reports, workplace decisions and your own explanations. It is deliberately slower than accepting a conclusion by tone. It is faster than repairing a decision built on a bridge nobody inspected.

Write the three parts in different sentences

Start with literal language. Do not improve the source's argument while transcribing it.

1. Claim: write the exact proposition being asserted. Make it specific enough to be wrong.
2. Evidence: list observations, measurements, records or source statements without adding the conclusion.
3. Inference: state why that evidence would make the claim more likely, including the assumptions required.

“This cable is future-proof” is a claim. “The listing says 240 W and USB 40 Gbps” is evidence only about what the listing says. The inference is that the listing is accurate, the exact shipped cable matches it, your devices support those modes and those capabilities cover your future use. Once written, the vague promise becomes several testable questions.

The seven-step inference audit

1. Narrow the claim

Remove words such as better, safe, proven, significant and works unless the source defines them. Identify the population, condition, outcome and time horizon. “The intervention works” becomes “in this study population, the intervention changed the prespecified outcome over twelve weeks compared with the control.” The narrower claim may still matter. It simply stops borrowing certainty from undefined language.

2. Inventory the evidence

Separate primary observations from summaries and repetitions. Ten articles repeating one press release are not ten independent observations. Record sample size, measurement method, comparison, missing data and source access when available. For a product claim, record the exact model, specification, certification source and manual—not a category-level description.

3. Name the bridge

Write the reasoning in an “if, and, then” form: if the measurement is valid, and the comparison controls the important alternatives, and the sample represents the intended case, then the evidence supports the claim to this degree. Each “and” is an assumption that can fail.

4. Generate rival explanations

Ask what else could produce the same evidence. Measurement error, selection, confounding, regression to the mean, incentives, ordinary variation and a different mechanism may fit. You do not need to prove every rival. You need to know whether the claimed explanation has actually outrun them.

5. Check independence

Independent evidence uses a genuinely separate route: another dataset, method, laboratory, source or prediction made before the result was known. A second analysis of the same records may test robustness, but it is not a new observation. A manufacturer quote copied by five stores remains one source.

6. Match confidence to consequences

A low-cost reversible choice can tolerate more uncertainty than a medical, financial, safety or public accusation. Increase the evidence burden when error is expensive, difficult to reverse or harmful to other people. The protocol is not a substitute for qualified professional advice; it helps reveal when that advice is warranted.

7. Write the calibrated conclusion

Use one of four outcomes: supported, provisionally supported, unresolved or contradicted. Add the condition that would change the judgment. This keeps uncertainty operational rather than ceremonial.

The one-page claim–evidence–inference worksheet

Three fast examples

An AI answer

The answer claims a regulation requires a particular action. The evidence offered is a citation. The inference assumes the source exists, is current, applies to the jurisdiction and says what the answer claims. Use the AI citation check before carrying the conclusion into real work.

A scientific headline

The headline claims a habit causes an outcome. The evidence is an observational association. The inference assumes important differences between groups were measured and controlled, the direction is correct and the sample generalizes. Read the methods and comparison before upgrading association to cause. The scientific-paper field guide supplies the deeper pass.

A product recommendation

The recommendation claims a device is best. The evidence is a specification table and several reviews. The inference assumes the use case matches yours, the reviews are independent, the compared models are current and the tested criteria matter more than omitted tradeoffs. Replace “best” with “best for this use under these constraints.”

What strong evidence does—and does not—do

The National Academies' Reproducibility and Replicability in Science report distinguishes computational reproducibility from obtaining consistent results in a new study. That distinction illustrates a larger rule: a second route can test a different part of the bridge. Re-running code checks whether the reported analysis can be obtained; collecting new data asks whether the result persists beyond the original sample.

NIST's AI Risk Management Framework emphasizes validity, reliability, transparency and context. Those properties are not a single score. A system can be accurate on one benchmark and unsuitable for a high-consequence use. The inference must include the intended context.

The Cochrane Handbook shows why evidence synthesis requires explicit criteria, bias assessment and attention to heterogeneity. Counting studies without examining how they were produced can make a large pile of evidence less informative than it looks.

Use a source ladder

When time is limited, check the closest available source first. A regulation outranks an article describing the regulation. A study report outranks a press release about the study. A product manual or certification database outranks a marketplace summary. Secondary sources can supply context and competing interpretation, but they should not silently replace the record that contains the relevant detail.

Record the source's role beside it: observation, method, interpretation, background or decision rule. This prevents a background explainer from being cited as if it produced the evidence, and it prevents an authority's general guidance from being stretched into a finding about one case.

Then run a subtraction test. Remove the most emotionally persuasive sentence, image or credential. Does the remaining evidence still support the claim? If the conclusion collapses when presentation is removed, the argument may be relying on authority, vividness or repetition rather than the stated bridge.

Use a stop rule

Verification can become endless. Before checking, define what would be enough for the decision. A reversible purchase might require a manual and return policy. A high-stakes claim may require primary records, independent corroboration and qualified review. The stop-rule protocol keeps rigor from becoming compulsive searching.

The purpose is not to eliminate inference. Evidence never interprets itself. The purpose is to make the inference visible enough to criticize, improve and revise.