Keep the disagreement in the record

Disagreement is not always contradiction

At 09:00, a local report says a border crossing is closed. At 11:00, an authority says it is open to commercial traffic. A third source repeats the first report without its timestamp. These records might describe a reopening, different categories of traveler, an inaccurate statement, or a genuine dispute. A system that immediately labels the second statement a correction has already made an analytical decision.

The first task is to preserve what each source said and the conditions under which the statements could conflict. W3C PROV-O supplies a vocabulary for entities, activities, agents, and derivation. It can help describe where a record came from; it does not tell us which source is right. That judgment needs a separate, attributable assessment. [1]

Four records, rather than one overwritten answer

Our proposed capsule design separates the passage, the extracted assertion, the relationship proposed between assertions, and the review decision. Extraction may be wrong even when the passage is authentic. A reviewer may accept the quotation while rejecting the interpretation attached to it.

This separation also prevents a status such as “reviewed” from becoming a substitute for the reason it was reviewed. A later analyst should be able to discover what evidence was available, which interpretation was accepted, and what would justify reopening the decision.

Do not ask formal consistency to settle a factual dispute

An ontology can specify what a relation means and which combinations of assertions are logically incompatible. That is different from establishing that either assertion accurately describes the world. OWL’s open-world semantics also matter: a missing statement is not automatically false. An application must decide explicitly when an incomplete record means “unknown,” rather than silently treating absence as a negative answer. [2]

A useful counterexample is two correctly represented but unreliable reports. They can satisfy a schema, have valid citations, and agree with each other because both copied the same account. Agreement within the data is not independent corroboration. The EEAS’s 2026 FIMI report similarly distinguishes observed activity sets from the actors to whom they may be attributed; a relationship in a network is not by itself an identity finding. [3]

When a graph earns its place

For a short case file, a table of claims and annotated documents may be enough. A relational database can also preserve assertions and review history. A graph becomes a useful option when recurring questions cross many relationships: which conclusions depend on this disputed source, which accounts share an origin, and which assessments must be revisited after a correction?

The retrieval contract matters as much as storage. A compact answer context should include the relevant assertion, its qualifications, material alternatives, and enough source text to inspect them. It should not retrieve only the currently preferred account. Removing duplicate reports may save tokens, but removing their lineage could conceal why ten apparent sources amount to one.

Evaluate the ability to change a mind

A proposed study would seed cases with time mismatches, copied reports, genuine contradictions, and legitimate differences in scope. Compare document retrieval, a claim table, and graph-assisted retrieval on the same questions. Measure missed counter-evidence, falsely identified contradictions, unsupported conclusions, and the time a second analyst needs to correct an assessment.

NIST’s citation probes distinguish faithfulness, completeness, and sufficiency. Those are useful checks for whether an answer cites accurately, omits important context, or overstates what evidence can establish. Report retrieval tokens, extraction cost, latency, and review effort alongside those measures. PRAXIS is the intended review workspace; DIALECTICA is the knowledge-engine boundary we are developing. No completed comparison or measured advantage is claimed here. [4]

A later account does not automatically win

C01 remains the earlier January proposal when C03 records July consideration. The Industry Forum’s transition request has its own attribution and does not become agreement merely because a newer source exists. A local reviewer can reject unsupported wording and preserve the correction without erasing older records. In a contested event, retain the allegation, denial and limits of observation separately; recency is not a mechanical truth rule. [5]