H-CRSPV: Preventing Semantic Omission in Late-Bound Large Language Model Releases
H-CRSPV rejects 36 LLM-release semantic-omission cases that per-object validation accepts.
The paper defines Semantic Omission under Valid Evidence, where per-object checks pass but a late-bound LLM release misses a required relation or binds evidence to the wrong object. Hybrid Cryptographic Relation-based Semantic Plan Verification commits an authority record, checks obligation coverage and evidence sharing, and finalizes the release only after a challenge window. Across 36 omission artifacts, object-level validation accepted all 36, while H-CRSPV rejected all 36 and accepted six honest releases. A prototype also rejected all 60 tested mutations and kept one authority record fixed across three legal Qwen2.5-1.5B releases.
- SOVE is omitting a required relation while each submitted object still verifies.
- H-CRSPV checks obligation coverage and one-to-one evidence binding before release.
- It rejected all 36 omission artifacts and accepted all six honest releases.
- A four-validator Qwen2.5-1.5B workflow kept one authority record across three legal releases.
From the original article
1 min read at arxiv.orgLarge-language-model release pipelines increasingly combine commitments, signatures, provenance records, and heterogeneous verification backends. Yet validating every submitted object does not establish that a release realizes every requirement of its registered transformation.
By Weijie Miao · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.