Publish the narrowest statement that the evidence supports. First check whether the sources describe the same thing, at the same time, in the same units. If a material contradiction remains, preserve it visibly instead of letting an agent choose whichever answer fits the draft.
This is a proposed editorial method, reviewed September 12, 2026. It is distinct from checking that a citation exists: the sources may all be real and still support incompatible claims. Worked numerical examples are illustrative, not reports of a vendor dispute.
- 01Normalize before adjudicating.A price, release or benchmark claim needs matching scope, date and units.
- 02Do not vote by citation count.Repeated coverage may come from one source, and two primary pages can disagree.
- 03Publish uncertainty precisely.State the supported fact and the unresolved boundary without inventing a reconciliation.
- 04Deliver a claim record, not just prose.An editor can approve a statement whose supported, contradicted and unresolved claims are listed; a polished paragraph hides which is which.
01 — Practical decisionWrite the disputed claim as one sentence
An agent may say that sources disagree about a product when the actual conflict is much smaller: one price tier, one access route or one launch date. Write the exact claim under review and list the evidence for each alternative. Keep neighboring facts out of the dispute until they are needed.
For a price, identify the product, plan, currency, billing unit, region and effective date. For a release, distinguish announcement, preview, general availability and catalog listing. For a result, identify the task, metric, evaluator and measurement conditions. Without that alignment, the agent is comparing labels rather than claims.
The source-independence guide helps trace repeated reporting to its origin. Independence is relevant here, but it is not the verdict. Two independent sources can describe different scopes accurately, while two pages from the same vendor can contain a real unresolved contradiction.
02 — MethodClassify the disagreement before choosing wording
The proposed decision table below separates apparent conflicts from unresolved ones. Use it as an editorial routing aid, not an automatic trust score. The same article may resolve one claim, qualify another and withhold a third.
Reuters’ standards put accuracy ahead of speed, emphasize honest attribution and call for clear correction of errors. Those principles support a slower, explicit answer when the evidence cannot sustain a clean headline. They do not mean that every uncertainty requires abandoning the whole article.
| Disagreement | Evidence needed | Publication choice |
|---|---|---|
| Different dates | Effective dates and revision history | Use the matching period; explain the change. |
| Different product surfaces | Endpoint, plan and access route | State each scope separately. |
| Different units | Billing or measurement definition | Convert only with justified assumptions. |
| Repeated secondary reports | Original source and derivation | Do not treat repetition as independent support. |
| Conflicting current primaries | Clarification or authoritative correction | Disclose the conflict or withhold the claim. |
| Insufficient evidence | A missing primary or observable result | Label unknown; narrow the conclusion. |
03 — The evidenceTrace who produced each version and when
Keep the source URL, title, publisher, retrieved date, visible document date and relevant passage together. Record whether the date refers to publication, revision or the event itself. A page fetched today can still describe an older price, and an undated live page may establish current availability without revealing launch history.
W3C’s PROV overview describes provenance through entities, activities and people involved in producing information. For editorial work, that means retaining the path from a claim to its source and through later transformations. Provenance helps inspect a claim; it does not make the claim true by itself.
For example, BFL’s FLUX Video Edit page establishes a currently presented product offering. In this research, that page did not establish a direct announcement date. The editing decision guide therefore uses current availability and leaves launch timing unasserted. That is a bounded negative finding, not evidence that the product has no announcement anywhere.
04 — The decisionWork through the units before inventing a compromise
Consider an illustrative conflict: a fictional summary says an edit costs $0.03 per output second, while another says $0.03 per source minute. For a ten-second output, the first would imply $0.30. The second cannot be substituted without verifying what input is charged and how partial minutes are handled. Averaging the two figures would create a price that neither source states.
The next action is to inspect the actual rate card and request contract, then ask whether the pages refer to the same endpoint and effective period. If one is demonstrably superseded, cite the current scope and record the correction. If both still claim authority and no explanation is available, publish the disagreement or omit the cost forecast.
Do not generalize this illustrative example into a current BFL pricing conflict: the BFL product page and endpoint documentation inspected for this batch agree on the output-second rate. Likewise, earlier discovery notes about other vendors are leads for a new investigation, not reusable proof of a current dispute.
05 — ExampleKeep the uncertainty beside the decision it affects
If an uncertainty changes the buying decision, it belongs next to the price or recommendation. A vague disclaimer at the bottom cannot repair a definite unsupported claim in a headline, table or description. Write what is known and which comparison remains unavailable.
Useful wording might say that a vendor page lists one rate while another current document describes a different billing basis, and that a written quote is needed before budgeting. Avoid presenting both as simultaneously applicable unless the sources explain how. Attribution is necessary but does not make every claim worth repeating.
The research-proof reference distinguishes the kind of evidence behind a result. Preserve that distinction during conflict resolution. A demonstration cannot settle an empirical benchmark dispute, and a vendor assertion is not independent validation just because the agent found a second page repeating it.
06 — The failure caseCorrect the whole artifact when the evidence changes
When the conflict is resolved, update the body, metadata, tables, captions, charts and downloads that depend on it. Keep a concise correction record naming the old claim, the new evidence and the affected scope. Do not silently leave the obsolete number in a worksheet that readers are likely to reuse.
If only one claim changed, retain the valid research around it. A focused correction should not create new unsourced claims or rewrite the article’s history. Record the actual review date, and distinguish that from the event date and the article date.
For AI transformation work, require the research agent to deliver a claim record along with the prose. That record should show supported, contradicted and unresolved claims. The human editor can then approve a precise statement rather than infer certainty from a polished paragraph.
As-of date. September 12, 2026: sources retrieved and reviewed. The article is dated September 12, 2026. Verified event dates are stated separately.
Sources and method. Original editorial decision method using Reuters standards and W3C provenance guidance. BFL availability inspected as a bounded real date example.
Limits. Sources checked September 12, 2026. No source-reliability score or current DeepSeek/Devin pricing adjudication; hypothetical rates are explicitly illustrative.
07 — The failuresFour reconciliations that invented a fact
An agent asked to resolve a disagreement will usually produce a resolution, and the resolution is where the invented facts live. The illustrative cases below each follow from skipping a step of the method. They use the same fictional pricing example as the units section and are not reports of a vendor dispute.
The averaged price
One summary says $0.03 per output second, another $0.03 per source minute, and the draft reports an average. Neither source states the resulting figure. The billing basis was never checked, so the compromise is a number that exists nowhere.
The newest wins
The page fetched most recently is treated as correct because it is newer. It describes a different plan on a different endpoint. Effective date, scope and authority were not compared, only the retrieval date.
The citation vote
Three pages repeat one figure and a single primary page states another, so the draft goes with the majority. The three are one source, syndicated. Repetition was counted as corroboration, which the table says not to do.
The footer disclaimer
The headline and table state a definite price while a line at the bottom says sources vary. The uncertainty changes the buying decision, so it belonged beside the price. A reader of the table never sees it.
Each case is a resolution chosen before the evidence was aligned. The claim record in the next section is the artifact that makes the alignment visible to the editor who has to approve the sentence.
08 — The methodThe claim record the agent should deliver
The correction section asks the research agent to deliver a claim record with the prose, and the record has a fixed shape. For each disputed claim: the claim written as one sentence, as the first section asks; each alternative version with its source URL, publisher, visible document date, retrieval date and the scope it describes; the classification from the decision table, from different dates and surfaces through repeated secondary reports to a real unresolved contradiction; the publication choice that classification implies; and the status the editor sees, supported, contradicted or unresolved, with the exact wording the draft is permitted to use.
An illustrative record for the fictional price claim reads: claim, the per-unit cost of a video edit on a named plan; alternative A, $0.03 per output second from a fictional product page dated this quarter; alternative B, $0.03 per source minute from a summary with no visible date, retrieved later; classification, different units with the billing basis unverified; publication choice, convert only with a justified assumption, otherwise state the two bases and request a written quote; status, unresolved; permitted wording, the vendor page lists one rate while another current document describes a different billing basis, and a written quote is needed before budgeting.
The record does three jobs at once. It shows the editor what was aligned before anything was adjudicated, so the four failures above cannot hide in a paragraph. It carries the provenance the tracing section asks for, so a later correction reaches every dependent claim. And it makes the article’s uncertainty precise, which is the conclusion’s point: a narrower supported answer, with its boundary stated, is publishable progress.
Do the sources describe the same thing, at the same time, in the same units? Until the answer is yes, there is nothing to adjudicate, only scopes to separate. The first three rows of the decision table end here.
09 — Next stepDisagreement between sources is a finding, not a blocker
Ask for the claim record on the next research task before asking for the prose. The record is short, it is where the editor’s decision actually happens, and it turns an agent’s confident reconciliation back into evidence someone can inspect.
A narrower supported answer is publishable progress
Resolve differences in scope and units where the evidence permits it. Where a contradiction remains, make it visible at the point it affects the reader’s decision. Preserve the source record so a later correction can improve the article without inventing certainty today.