llm-eval-ci-gate-rtl-verilog

Quorum is a dual-tool CI evaluation gate for LLM-generated RTL, built in Python, Docker and GitHub Actions. Its rule is that no result counts on one witness: for anything a tool can check, two structurally different tools must agree, and for anything only a human can check, two independent reviewers apply the same written rubric, with the agreement rate itself measured rather than assumed. The gate runs static analysis and simulation, with golden sets, tool-grounded scoring, pre-registered pass bars, and CI quality gates on prompt, model and retrieval changes.

Measured results are the point of the repo: 27 single-defect mutants seeded into six known-good reference modules, run through both lanes. The gate passes 6 of 6 known-good modules (no false positives), kills 22 of 27 mutants (81.5 percent), and the lane breakdown is the thesis: 19 mutants were caught by the dynamic lane (Icarus) only, 1 by the static lane (Yosys) only, 2 by both, and 5 survived both, all five being equivalent mutants. A static-only gate would have shipped all 19 dynamic-only catches.