Remote work is permitted on Fridays.
The passage supports this claim.
A source can be present.
And still contradict you.
Lexical overlap is not enough to establish evidential support. A failed baseline makes that distinction concrete.
Claim support
Read the relationship. Then read the score.
The lexical baseline failed: contradictory examples scored higher on average.
Synthetic reading exampleA citation can
A citation can
contradict the claim.
Claim: “The policy allows remote work on Fridays.”
Reading examples are synthetic, not recovered experimental rows. Reported baseline AUC: 0.008 against 0.65. The separate contradiction-detector evaluation is not the live gamma implementation.
The K3 result
A failed baseline stays in the record.
| Reported measure | Value |
|---|---|
| Supported examples · mean score | 29 examples · 0.780 |
| Contradicted examples · mean score | 64 examples · 0.890 |
| Silent examples | 12 examples |
| AUC / threshold | 0.008 / 0.65 |
Contradicted examples scored higher on average than supported examples. This baseline cannot be presented as validated evidential support.
Separate detector
Promising numbers need their error rates.
A subsequent deterministic detector was evaluated on a 95-example holdout. Reported contradiction recall was 88.1%, with an 11.5% false-positive rate on supported examples and 0% on silent examples.
These subgroup error rates belong with the recall figure. P-random calibration and integration remain open. This detector is not the live gamma support score.
Product implication
Show method and status beside the value.
A source reference, retrieval score and claim-support assessment answer different questions. The interface should preserve that distinction.