Self-Consistency, Not Independent Verification: An Independent Audit of What "PROVEN" Means in EquivSVA

A PROVEN verdict from a formal engine is commonly read as “this assertion independently verifiesdesign intent.” We audit the public AI-assertion dataset EquivSVA and report three findings: (1)in 80 of 120 behavioral families, gold properties and RTL are generated from the same spec.json,and the generator reproduces them byte-for-byte; the remaining 40 are not covered by this conclu-sion. (2) 131 invariants receive no reachability witness, while the generator’s own comment stillclaims “Reachability/non-vacuity witnesses.” (3) Of 93 formally sound properties, 79 (85%) are notdistinguished by any of 3 controlled mutants, with UNKNOWN = 0 ruling out non-convergence.We therefore read PROVEN as “self-consistency of two co-generated artifacts,” not “independentcharacterization of intent.” We do not claim EquivSVA is defective, nor that its formal resultsare wrong; we bound the reading “PROVEN = independent verification” along source, coverage, andfailure-mode axes. A minimal reproducibility package reproduces at least three core findings onpublic data.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-02
DOI
https://doi.org/10.5281/zenodo.23891926
Primary Topic
Formal Methods in Verification
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
preprint

Self-Consistency, Not Independent Verification: An Independent Audit of What "PROVEN" Means in EquivSVA

zhizhi(廖至智) liao
Zenodo (CERN European Organization for Nuclear Research)
Formal Methods in Verification
preprint

Self-Consistency, Not Independent Verification: An Independent Audit of What "PROVEN" Means in EquivSVA

zhizhi(廖至智) liao
preprint en

Abstract

A PROVEN verdict from a formal engine is commonly read as “this assertion independently verifiesdesign intent.” We audit the public AI-assertion dataset EquivSVA and report three findings: (1)in 80 of 120 behavioral families, gold properties and RTL are generated from the same spec.json,and the generator reproduces them byte-for-byte; the remaining 40 are not covered by this conclu-sion. (2) 131 invariants receive no reachability witness, while the generator’s own comment stillclaims “Reachability/non-vacuity witnesses.” (3) Of 93 formally sound properties, 79 (85%) are notdistinguished by any of 3 controlled mutants, with UNKNOWN = 0 ruling out non-convergence.We therefore read PROVEN as “self-consistency of two co-generated artifacts,” not “independentcharacterization of intent.” We do not claim EquivSVA is defective, nor that its formal resultsare wrong; we bound the reading “PROVEN = independent verification” along source, coverage, andfailure-mode axes. A minimal reproducibility package reproduces at least three core findings onpublic data.

Zenodo (CERN European Organization for Nuclear Research)
Formal Methods in Verification
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon