Settledness and contradiction

This page covers the axis that says how settled a claim is, the deterministic check that keeps a fact from stating more than its own sentence did, and the path by which a later document retracts an earlier claim. It follows Grounding and consolidation, which owns the rest of the audit and the cascade.

Two independent axes

EpistemicKind answers whose claim this is. Stance answers how settled it is. Folding them together would make every combination a new enum member, so they stay apart and a claim carries both.

  settled  ->  reported  ->  hedged  ->  disputed  ->  refuted
  plain        credited      qualified   sources      the source
  assertion    to somebody   or doubted  disagree     says it is wrong

The ladder only ever runs downward. at_least takes the less settled of two readings, because evidence about a claim may lower confidence in it and never raise it. Two properties do the work elsewhere. distorting, true from hedged down, decides whether reading a claim flatly would change what it says. decisive, true for settled and refuted, decides whether a claim is definite enough to close one it contradicts.

Only an unsettled stance is stored, in the claim’s attributes beside the speaker and the epistemic kind, so nothing is written for the ordinary case and no schema changed.

Why a quote check was not enough

A hedged sentence contains a contiguous, character-exact substring that asserts its confident half alone.

  source    RIRobustNetV7 improves over RIConv++ on noise, though the
            margin falls within run-to-run variance.
  quote     RIRobustNetV7 improves over RIConv++ on noise
  fact      RIRobustNetV7 beats RIConv++ on noise and occlusion.

Every check that existed passed. The quote is verbatim, contiguous and locatable, and the stored fact says the opposite of what the author wrote. The system behaved as specified and the specification was wrong for hedged content.

Certainty is a comparison

Qualification.read expands the located quote to the sentence around it and reads both that sentence and the fact’s own statement through one certainty detector, then acts on the difference.

The statement is the compared side rather than the quote, because the statement is what gets stored, embedded, ranked and handed to a reader. A quote spanning a qualifier the statement then drops still leaves a flat assertion in the graph. The detector maps four marker registers onto Stance, and its lexicon is a cheap detector rather than the decision, since a register it does not know is missed symmetrically on both sides and costs an error in neither direction.

Two differences reject the fact as stripped_qualifier. A statement less doubtful than a sentence that is itself at hedged or worse, and a negation in the sentence the statement drops. A dropped attribution does not reject, because the claim survives intact and the stance plus the paired source excerpt carry what was lost. Whatever survives is stamped through settledness, which may only lower what the model proposed.

What the reader gets

LiveFact.line renders the marks in one bracket, and a settled world claim with no speaker prints none.

  - [hedged] (improves_over) V7 improves over the baseline on noise
  - [Maya, opinion, disputed] (uses) the batch lane is the wrong default

The label is not the whole defense. Published 2026 work finds the evidential register barely discounted by downstream readers, so an unsettled claim also changes what recall returns. deduplicate stops treating a source excerpt as a repetition of an unsettled fact from the same span, which is what makes the source-wins rule in Sources and derived knowledge a property of the answer rather than advice to the reader. The rendered result also opens with a standing instruction naming the excerpt as the authority.

Closing a claim a later document disproves

A source replaced in place already closes everything built from its old text. A later, different document showing an earlier claim was wrong had nothing to close it, so the claim stayed live and kept answering questions.

TimedFact.contests routes that case straight to the batched model call, skipping the deterministic tiers, because a correction reads almost exactly like the claim it corrects and similarity would call it a near-duplicate. It is true when the supporting sentence announces a correction or when extraction read the source as disputing or refuting the claim.

The model may then answer REFUTE, naming a claim from that candidate’s own catalog and no other. GraphWriter._settle_contradictions applies it conservatively.

The new fact’s stance	What happens to the claim it contradicts
settled or refuted	closed through Fact.Claim.refute, stamped with the chunk that closed it
anything less certain	left live and marked disputed, and the new claim joins the standoff

A false retraction silently deletes something a person committed to memory while an unresolved disagreement is merely visible, so the asymmetry is deliberate.

This is also the correction path a caller has. There is no tool that edits a derived claim, because derived claims are read out of sources and any hand edit would be overwritten. A keep note saying plainly that a claim is refuted goes through this same path and closes it, with the note recorded as the evidence that closed it.

Measured, not asserted

chefe run aizk-eval certainty scores this on a committed corpus where every quote is a contiguous, character-exact substring, so each case already passes quote verification. The before arm is the same audit without the comparison, which is what the code did when the defect was reported.

Measure	Before	After
flattening admitted, over 12 distorting proposals	1.000	0.083
false rejection, over 20 faithful proposals	0.000	0.050
source excerpt paired with an unsettled claim	0.000	1.000
settledness read matches the human label		0.969

Extraction and models records the lexicon ablation behind those numbers and reads the two remaining costs.

Next

• Grounding and consolidation owns the rest of the audit.
• The bi-temporal model explains the ranges a refutation closes.
• Evidence and provenance is the reader-facing view.