Skip to content

Controlled structural-validity result

Run date: 2026-07-02

Toolchain:

  • SV-Gap reference oracle 0.1
  • Yosys 0.66
  • Icarus Verilog 13.0
Witness Functional unsafe Structural unsafe Functional safe Structural safe
Stable level crossing pass fail (REF-CDC-001) pass pass
Combinational source crossing pass fail (REF-CDC-002) pass pass
Multi-bit counter crossing pass fail (REF-CDC-003) pass pass
Asynchronous reset release pass fail (REF-RDC-001) pass pass

Across the eight deliberately balanced candidates, all eight pass their functional simulation and four fail the configured structural oracle. The CLI therefore reports a structural-validity gap of 0.500.

Interpretation: 0.500 is a harness-validation result created by the paired experimental design. It is not an estimate of defect prevalence in models, benchmarks, or production RTL. The defensible conclusion is narrower: for each controlled witness, the functional oracle assigns the same successful outcome to candidates that the declared structural oracle distinguishes.

What a researcher can take from this

The useful result is not the balanced fraction. It is the observable failure of a functional pass to settle a declared production question. The evidence profile makes that handoff explicit: what the functional oracle answered, what the configured structural rule rejected, and what evidence should be added next. A model or benchmark team can apply the same representation without claiming that this small controlled suite estimates real-world defect rates.

Continue with one of these paths:

Reproduce the evidence

The pinned container reproduces the asynchronous-reset witness. The first image pull is large and depends on network speed; execution is quick once the image is cached:

docker run --rm ghcr.io/shsridhar-beep/svgap:v0.3.0-alpha.8 demo

To regenerate all four controlled witness pairs from a source checkout after installing the documented open-source prerequisites:

python3 -m unittest discover -s tests -v

Candidate-level machine-readable reports are regenerated under each example's build/ directory and intentionally excluded from source control.