Tristan Lim and colleagues submitted Pre-QED: Environment Specifications for Agentic Verification to arXiv on 11 November 2026. The authors argue that the bottleneck in verification using large language model agents has shifted from producing proofs to specifying what those proofs should cover. They propose executable verification harnesses intended to model all legal uses of an application programming interface (API).[1]
The abstract criticizes approaches that reuse fuzz-testing harnesses as formal specifications. It names OSS-Verify from Anthropic and Theorem, saying its agents prove that individual harnesses never crash. In the authors’ account, a fuzz harness follows one sequence of API calls in one calling context, so proving that harness covers a usage scenario rather than the whole library. This is the preprint’s characterization; our reading does not independently establish OSS-Verify’s coverage.[1]
Pre-QED’s proposed workflow extracts an API’s usage contract from code and documentation, then uses Linear-time Temporal Logic (LTL) to exclude illegal call orders. The authors say they release their harnesses as open-source challenges and identify difficulties in proving them. Coverage of every well-behaved client is the proposed objective, not a demonstrated result established by the saved abstract.[1]
Sources
Check or correct this page
Verification record
The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.
682064cdf22c37f1aa22cc3a38ec7c291800bf8b5ef0d26548a8d1bd415adc24