Research Evidence Verifier

Evaluate whether a claim is actually supported by current, primary evidence while separating fact, inference, uncertainty, and contradiction.

The result you're building

A claim ledger that states what is supported, contradicted, uncertain, or inferred; links each conclusion to current primary evidence; records dates and scope; and gives a reproducible method for updating the answer.

Use this guide when

• A decision depends on whether a factual claim is actually true now.
• Sources disagree, repeat one another, or use statistics without definitions.
• An AI-generated answer needs independent verification and provenance.

Do not use it as a substitute for

• Manufacturing certainty when the available evidence is missing, proprietary, or inherently predictive.
• Treating source count as quality when many pages copy the same original claim.

Before you change anything

• Collect the items below first. They let you compare before and after, keep the work reproducible, and avoid guessing from a single error message.
• One atomic claim written with subject, predicate, scope, date, and measurable terms.
• Decision context and accuracy/freshness threshold.
• Candidate primary sources and the exact records, methods, or statements they provide.
• Known definitions, units, population, jurisdiction, and version.
• A place to record citations, access dates, quotes/paraphrases, and inference.

Understand the system before fixing it

Atomic claims are testable; bundled claims are not
Split 'safe, profitable, and legal' into separate claims with different evidence standards. A source supporting one adjective does not support the others.

Primary does not automatically mean unbiased or complete
Official data may be authoritative for its own policy or records but still omit failures, use narrow definitions, or have incentives. Evaluate fitness for the claim.

Fact, interpretation, and prediction need separate labels
Evidence may show what happened; a causal explanation or forecast is an inference unless the study design directly supports it.

Date belongs inside the claim
Software behavior, prices, laws, officials, and product policies change. Record observed/published/effective dates and when to recheck.

Evidence-to-decision map

Step-by-step procedure

Work in order. Record the output after each step. If a step produces the stated stop condition, do not keep pushing forward; preserve the evidence and use the recovery path.

Step 01 — Rewrite the question as atomic claims

Why: A broad question lets weak evidence appear to answer more than it does.

Do: Break it into individual statements with who/what, measurable verb, scope, time, and definitions. Mark whether each is factual, causal, normative, or predictive.

Read the result: If two clauses could have different truth values, split them.

Next: Assign an evidence requirement to each claim.

Step 02 — Define source and freshness standards

Why: A source can be credible yet unsuitable for the specific decision.

Do: Rank records: direct official data/specification, primary research/filing, independent measurement, expert synthesis, reporting, and unsupported commentary. Set a maximum acceptable age.

Read the result: Use source ownership, methodology, access to underlying data, and incentives to grade fitness.

Next: Search for disconfirming evidence as deliberately as confirming evidence.

Step 03 — Extract evidence with scope attached

Why: Quotes and numbers lose meaning when separated from definitions and context.

Do: Record source, publisher, author/owner, publication/effective/observation dates, exact locator, claim supported, units, population, limitations, and a short compliant excerpt or paraphrase.

Read the result: If you cannot state which exact claim a source supports, do not cite it generally.

Next: Separate raw evidence from your interpretation.

Step 04 — Build a contradiction table

Why: Disagreement often comes from different definitions or periods rather than one source being false.

Do: Put conflicting values side by side with methods, units, sample, dates, and scope. Trace shared upstream sources.

Read the result: Resolve only when one source directly fits the claim better; otherwise report the range or unresolved contradiction.

Next: Reduce confidence when independence is weak.

Step 05 — Issue a bounded conclusion

Why: Decision usefulness comes from knowing what is proven and what remains unknown.

Do: For each atomic claim label supported, contradicted, mixed, insufficient, or not verifiable; add confidence and why. Mark every inference explicitly.

Read the result: No citation may support more scope than its content.

Next: Add recheck date/trigger and missing evidence request.

Step 06 — Make the result reproducible

Why: Research decays if the next reviewer cannot rerun it.

Do: Save search terms, source list, access dates, exclusions, calculations, and claim ledger in a stable format. Version the conclusion.

Read the result: A reviewer should be able to reach the same classification from the recorded evidence.

Next: Schedule updates for unstable claims.

Worked example

Evidence collected

• Vendor case study reports 60% for one customer but does not disclose baseline ticket mix.
• Customer press release confirms deployment, not the cost figure methodology.
• No audited cost data or independent study is available.
• The metric excludes implementation and review labor.

Decision: The attributable claim 'vendor reports 60% in one case study' is supported. The general claim that the product reduces a typical buyer's total support cost by 60% is not established.

Actions taken

• Split reported case result from general effectiveness claim.
• Recorded metric definition, omitted costs, population, and source relationship.
• Requested baseline, total-cost method, sample size, and independent validation.

Why this example matters: Truthful research sometimes produces a narrower statement than the buyer wants. That constraint prevents marketing evidence from becoming a false guarantee.

Verify, recover, and hand off

Completion tests

• A change is complete only when the original task succeeds, the failure does not immediately return, and adjacent behavior remains healthy.
• Every conclusion maps to one or more exact evidence records.
• Dates, definitions, units, population, and scope are visible.
• Derivative sources are not counted as independent confirmation.
• Contradictions and missing evidence are stated, not hidden.
• Facts, inference, opinion, and prediction have distinct labels.
• Recheck trigger matches the claim's volatility.

Rollback or safe recovery

• Withdraw or downgrade a conclusion when its source changes, is retracted, or no longer matches scope.
• Restore the prior version in the ledger rather than overwriting history.
• Notify downstream users when a material claim classification changes.

If the expected result does not appear

Reusable handoff record

• Save this with the project, ticket, or client delivery. It turns the work into a repeatable result instead of a one-time guess.
• Atomic claim list with types, scope, and decision threshold.
• Evidence ledger with source lineage, dates, locators, and limitations.
• Contradiction and independence analysis.
• Per-claim status, confidence, inference label, and rationale.
• Version, search method, exclusions, and recheck trigger.

Agent delivery contract

Required inputs

Returned output

Agent refusal and escalation rules

• Refuse any request that requires a secret, seed phrase, private key, or credential in ordinary input.
• Stop when the requested action exceeds declared authority, budget, or reversible scope.
• Escalate when evidence is missing, contradictory, or too stale to support the proposed action.

Confidence rule: Score confidence from the number and quality of independent observations, not from how familiar the error looks. Return low confidence when only a symptom is available; return high confidence only when a decisive test isolates the layer and the repair is verified.

Official reference starting points

• https://www.cochrane.org/about-us/evidence-based-health-care
• https://www.nist.gov/topics/metrology
• https://www.loc.gov/item/lcwaN0009954/