Assumptions
Everything the model believes, stated before the math. If you do not buy one of these, do not buy the output.
- Detection is ingest-side and deterministic: two claims on the same (entity, field) differing beyond a fixed tolerance pair into the conflicts collection as first-class objects.
- The AI never silently resolves a conflict. Both sides stay visible until primary evidence settles it.
- Tolerances are fixed policy per field family, versioned.
Derivation
Pairing rule: claims c1, c2 on the same (entity, field) conflict when their relative gap exceeds the field family's tolerance tau:
At ask time the model prices the disagreement: the magnitude of the gap, the evidence weight on each side (M08), and which downstream outputs move if each side is believed - a leave-one-side-out recomputation of the affected models:
Identification strategy
Why the estimate supports the claimed meaning - and where the model refuses to claim more than the data carries.
- A conflict is an object, not a problem to be cleaned: keeping both sides on the record is what makes the disagreement auditable and prevents silent data cleaning from laundering a judgment call into a fact.
- The leave-one-side-out consequence is identified interventionally: the number shown is the actual recomputed difference in the affected model outputs, not a narrative about materiality.
Inputs and outputs
Reads
- Conflict data
- Spec Section 32 (collection 33): paired claims with tolerances
- Evidence graph
- Weights per side from Evidence Weighting (M08)
Writes
- Flagged contradictions with both sides shown, their evidence, and the consequence of each side being right
Calibration and gate evidence
Validation design
- Tolerance thresholds tuned on known contradictions in the resolved record; false-pair rate reviewed on ingestion samples.
Evidence on record
- False-pair review results are not yet published in the sealed source.
Failure modes
The ways this model declines, degrades or goes UNKNOWN - and what it prints when it does.
- Fields without a defined tolerance family do not pair; near-miss claims stay separate rather than being force-matched.
- A conflict whose both sides are stale stays open and reads stale - time alone never resolves it.
Sensitivity
- Tolerance choice trades false pairs against missed conflicts; the versioned table is the dial, and changes are reviewed against the false-pair rate.
Iron rules
The AI never silently resolves a conflict. Both sides stay visible until primary evidence settles it.
Source and audit
Written from the sealed model source. Every run is recorded - model version, frozen inputs, input hash, timestamp. There is no unrecorded model execution.