Two records appear to contradict each other
The coherence sweep is a model pass over the organisation’s written notes, run on demand by the steward. It compares each note with the structural records as they stood when the sweep began, and with other notes on the same team or page. Structure that lands while the sweep runs waits for the next sweep.
Each proposed contradiction is judged twice, by two independently framed questions: can both statements hold at once for the same subject, scope and time? and do they assert different facts about the same thing?
- When both answers agree that they cannot both hold, the findingFindingA difference in the records that needs a decision. It includes its sources and can describe a slow handoff, disputed fact or missing owner. is confirmed and quotes both sides.
- When the two answers disagree, or either cannot decide, the finding is marked UNRESOLVED and says so in its first words. That is a review item for a person, not a conclusion.
- When both agree that the statements can hold together, nothing is raised.
A contradiction is a judgement, so the product never clears one on its own. Resolve it by re-checking one side: retract the wrong note, correct the map, then resolve the finding with a note.
Two limits to hold. The pass is a model’s judgement and can miss a contradiction or propose one that does not stand; the release gates check it on fixed pairs, which raises the odds but does not make it certain. And a sweep that raises nothing is not proof that the notes agree. An omission is never a contradiction: the sweep does not flag a note for what it leaves out.
A sweep can also fail to evaluate a comparison, when the model gives no usable answer even after a second try. Such a comparison is counted as unevaluated in the sweep’s report, and the report’s status reads partial or failed instead of complete. When the model’s reply carried usable proposals beside items it could not read, the usable ones are judged and the comparison is marked partial with the count of discarded items. Nothing is invented in either case. A partial sweep is not a clean one: run it again, and if the same comparisons stay partial or unevaluated, tell the operator.