Two records use an identical term but define it differently. “Active customer” may mean a login within 30 days, a paid subscription at period end or any account with an open contract. The values cannot be compared until the operational definitions are aligned. The safe resolution is usually to rename and preserve both constructs rather than forcing one definition onto the other.

The observations refer to different countries, devices, market segments, query variants, account types or inclusion rules. A value for enterprise customers does not contradict a value for all users. Partition the claim by population and denominator. Only generalize beyond those boundaries when an additional, justified inference supports the extension.

A status, ranking, price, policy or feature can change. Records from T₀ and T₁ form a sequence if each accurately captures its own time. Preserve both timestamps, establish the order of states and decide which record is valid for the present claim. Never treat “latest” as equivalent to “most reliable” without checking capture quality and effective date.

Percent, percentage points, counts, rates, shares and index values are not interchangeable. Even equal units may hide different denominators. Convert only when the mathematical relationship is valid and every required quantity is known. If conversion is impossible, report distinct measures and explain what each one represents.

A survey, direct capture, model estimate and administrative record can observe different aspects of the same subject. Method divergence requires inspection of sampling, instrument behavior, missingness, transformations and error paths. Prestige cannot replace this inspection. The method that best fits the exact claim may be more useful than the method with the strongest general reputation.

When entity, proposition, definition, scope, time, unit and method are sufficiently aligned, disagreement may be genuine. Preserve the conflict, lower confidence where appropriate and identify a discriminating observation. “Unresolved” is not a failure when the available evidence cannot justify a winner; it is the accurate state of knowledge.

Preserve the source wording, then record each normalized subject, relation, value, qualifier and negation. Normalization must clarify meaning without silently strengthening or weakening the original assertion.

Identify the originating evidence line, capture date, version, custody and every transformation used in comparison. Derived summaries remain linked to their preserved inputs.

List identity, definition, scope, time, unit, method and lineage. Mark each axis as aligned, repairable, incompatible or unknown, with a short justification.

Record conversions, partitions, temporal versions, exclusions and weighting decisions. Include the rule and inputs so another analyst can reproduce the operation.

Use an explicit state. “Resolved” means the contradiction has been explained under the declared boundary; it does not mean that every related uncertainty has disappeared.

State remaining uncertainty, current confidence and the evidence that would reopen the conflict. A later source, changed state or better method may require revision without making the earlier process irrational.

Compare each normalized proposition with its source record. Check whether qualifiers, uncertainty, negation or population limits disappeared during rewriting. A clearer sentence is not valid if it changes what the source asserted. Keep both the original wording and the analytical form so semantic loss remains detectable.

Verify aliases, subsidiaries, predecessors, products, domains and versions. Two sources may refer to related but non-identical objects. A company statement cannot automatically be assigned to every product, and a historical domain state cannot automatically describe the present organization. Record the identity link and its confidence.

Compare the population, geography, channel, device, market and time of the final claim with every supporting record. A narrow observation may support a narrow conclusion while failing to justify a universal one. If the conclusion extends beyond the observed boundary, label the extension as inference and explain its basis.

Check formulas, exchange rates, unit conversions, deduplication, category mapping and aggregation. Inputs, versions and rounding rules must be available. A transformation can make evidence comparable, but it also introduces assumptions and error. If the result changes materially under a reasonable alternative rule, report that sensitivity.

Trace citations and extracted values upstream. Group records that originate from the same dataset, announcement, study or observation. Multiple distribution paths improve discoverability, not independence. When a conflict is decided by apparent majority, confirm that the majority does not collapse into one copied source.

Ask why one evidence line received more weight. The answer should identify better access, temporal fit, directness, method fit, independence or lower known error. “More authoritative,” “more popular” or “more precise” is insufficient without explaining how that property improves support for this exact proposition.

Resolution may explain why two records differ without establishing the true value of the underlying construct. Separate removal of contradiction from proof of a final answer. List unobserved populations, unavailable denominators, inaccessible methods and alternative explanations that survive the repair. Confidence should reflect those remaining gaps.

Define a date, state change, better measurement, new primary record or failed prediction that would trigger review. Without a reopening condition, a resolution can harden into dogma. With one, the conclusion remains stable enough to use while still responding rationally to stronger or more current evidence.

After definitions, units and boundaries are aligned, the values fall within the expected uncertainty range. Retain both evidence lines and report convergence. Do not describe harmless measurement variation as contradiction.

Each record is valid for a different population, period, version or construct. Rewrite the result as multiple bounded propositions. Partitioning resolves the apparent contradiction without pretending that the underlying states are identical.

The older record remains historical evidence, while a newer verified observation governs the present-state claim. Preserve the transition date and effective boundary. Supersession changes applicability, not necessarily the truth of the earlier capture.

Authenticity failure, corruption, invalid measurement, wrong identity or decisive methodological error can disqualify an evidence line for the claim. Keep the rejected record, defect and rejection rule so the decision can be audited.

Aligned, credible evidence continues to disagree, or the decisive coordinates remain unavailable. Report both lines, reduce confidence and define a discriminating test or review trigger. Retained conflict is more useful than manufactured consensus.

When action cannot wait for complete resolution, state the assumption that selects the working claim, its downside if wrong and the evidence that would reverse the position. Conditional use must never be presented as unconditional proof.