VERITYGATE
Paper at the GroundLM 2026 workshop, co-located with EMNLP 2026, on checking whether a model's written explanation actually follows the structured evidence it cites.
Accepted at GroundLM 2026, the workshop on Grounding Language Models: Learning Faithfully and Efficiently, co-located with EMNLP 2026.
A model can write a fluent explanation of a number without that explanation following the record it came from. VERITYGATE is a checker that tests a narration against a fixed schema in four gates: the evidence identifiers it declares, the entities it names, the numbers it states, and the type of claim it makes. It is deliberately narrow. It verifies the contract, not every fact in the prose, and the paper is careful to call its results verifier rejection rates rather than hallucination rates.
