Open Brain
arXivPaperNeeds Review

Verify the Verifier: Specification-Grounded LLM Cross-Checking for Bugs That Escape UVM Scoreboards

A framework checks passing hardware regressions against specifications to find bugs missed by UVM scoreboards. In tests across four OpenTitan IPs, it confirmed 41 specification violations among 68 candidate escapes.

arXiv · Nov 19, 2026
Open Source Open MarkdownOpen JSON
Source Summary

The framework checks traces from a passing UVM regression against the specification. Across **564 inserted bugs** in **four OpenTitan IPs**, it identified **68** candidate scoreboard escapes and confirmed **41** with assertions that failed on buggy designs and were proven on the original.

Practical Implication

For agent-assisted verification, a green regression only establishes agreement with what the scoreboard checks. Add specification-grounded checks when correctness must cover behavior the reference model may omit.

Agent-Ready Context
The framework checks traces from a passing UVM regression against the specification. Across **564 inserted bugs** in **four OpenTitan IPs**, it identified **68** candidate scoreboard escapes and confirmed **41** with assertions that failed on buggy designs and were proven on the original.

For agent-assisted verification, a green regression only establishes agreement with what the scoreboard checks. Add specification-grounded checks when correctness must cover behavior the reference model may omit.

These results come from inserted bugs in OpenTitan IPs. The material does not establish how often the method finds escapes in other designs or test environments.
Connected Context · Feed7 Judgment

This study shows that a passing UVM regression can miss behaviors absent from its scoreboard, and tests a way to derive additional checks from the specification. Of 564 inserted bugs, it confirmed 41 scoreboard escapes with assertions that failed on buggy designs and were proven on the original. The evidence supports adding specification-grounded checks in the tested OpenTitan setting; broader coverage remains unestablished.

SWE-Gate: Passing Functional Tests Is Not Enough for Software Engineering AgentsBoth show that passing functional tests can leave important requirements unchecked; SWE-Gate uses review-derived constraints, while this study grounds additional checks in hardware specifications.SWE-Serve: Benchmarking Agentic Engineering For Production Inference ServingBoth add an acceptance gate beyond local checks: this study checks specification behavior in hardware traces, while SWE-Serve tests production-style end-to-end inference behavior.EquivSVA: A Formally Verified Dataset of Behavioral Assertions Across Equivalent RTL ImplementationsBoth use behavioral assertions to test requirements beyond one implementation’s existing checks; EquivSVA additionally tests whether assertions hold across equivalent RTL implementations.
Context Map
benchmarkcoding#agent-evals#benchmark-integrity
Uncertainty
These results come from inserted bugs in OpenTitan IPs. The material does not establish how often the method finds escapes in other designs or test environments.