Datasets:
P1–P5 probe construction
Probes were handcrafted from each base conditional relation tuple. They were not generated by a model, so there is no probe-generation prompt or generator to run. Each tuple family has five intended claims:
| Probe | Transformation | Gold label |
|---|---|---|
| P1 | Contradict the relation | Not Supported |
| P2 | Replace the condition | Not Enough Evidence |
| P3 | Remove the condition | Not Enough Evidence |
| P4 | Preserve the original condition-scoped relation | Supported |
| P5 | Contradict the condition | Not Supported |
Boundary rule: the label concerns whether the complete claim is licensed in the
given paragraph. General plausibility outside the supplied text is irrelevant.
Tuple_ID is the family key and ProbeID identifies the individual claim. The
canonical files in data/probes/ contain the released claim text and label.
There was no distinct machine-readable “manual filtering and audit decisions” file.
The deposited canonical rows reflect the completed manual construction. The review
spreadsheets retained in provenance/manual_review_sources/ are contextual source
materials and should not be described as an exhaustive decision ledger.