Comprehensive independent RTL designer review protocol
Status: frozen before external review.
Reviewer packet SHA-256:
df961c824be3f7a48ec22312e7449a9ef57b6e523107d5b6a67197925042f83e.
The packet itself remains private until first-pass labels are locked.
Purpose
This is the project's single comprehensive human-expert review request. It evaluates whether SV-Gap's implemented evidence contract helps close a research-to-production RTL gap. It does not estimate a universal defect rate or rank models.
Functional benchmark acceptance does not establish every production-relevant
property. CDC, reset release, multi-bit transfer, and power-on policy require
declared intent and distinct evidence. SV-Gap preserves functional outcomes,
reports configured structural questions separately, and distinguishes pass,
fail, unknown, and tool_error.
The reviewer assesses technical correctness, intent sufficiency, safe claim boundaries, evidence usefulness, appropriate abstention, and missing pre-deployment evidence.
Review unit and timebox
The packet is designed for one to two hours and contains:
- five controlled safe/unsafe witness pairs covering every implemented primary
rule family:
REF-CDC-001,REF-CDC-002,REF-CDC-003,REF-RDC-001, andREF-XPROP-001; - three blinded generated candidates exercising reset release and the broader
selective-reset shapes represented by
REF-XPROP-002andREF-XPROP-003; and - a repository-level production-risk questionnaire.
This diagnostic selection provides breadth across the repository without turning expert review into full-study annotation. It is deliberately not a representative sample.
Judgments
For each controlled pair, the reviewer records whether the pair supports the stated rule distinction, does not support it, or is uncertain, with technical evidence and missing qualifications.
For each blinded candidate, the reviewer records pass, violation, or
uncertain, confidence, source-level evidence, and notes. The reviewer judges
the supplied specification and RTL, not model quality or likely silicon
failure.
The repository-level review asks whether:
- the production questions are legitimate and correctly scoped;
- the manifest carries sufficient intent;
- the evaluator abstains in the right situations;
- clean and failing results have safe meanings;
- evidence is actionable for an RTL team;
- key checker failure modes are disclosed;
- important deployment evidence remains explicit; and
- the evidence profile improves the research-to-production handoff.
Independence and lock
Model identity, sample identity, functional outcomes, automated candidate outcomes, and private mappings are withheld. Before returning first-pass answers, the reviewer is instructed not to consult the repository, automated reports, source-text searches, or other labels. Prior exposure and tool use are disclosed.
Returned files are timestamped, content-hashed, and retained unchanged before automated comparison or reconciliation. Later discussion or revised judgments remain separate from the original review.
Permitted claims
The review may support statements that one independent RTL designer:
- confirmed, rejected, or qualified specific rule constructions;
- assessed representative applications of the rules;
- identified missing intent, evidence, and abstention conditions;
- evaluated whether the profile maps to production review practice; and
- recommended concrete technical or methodological changes.
It does not support:
- defect prevalence or model-quality estimates;
- population-level false-positive or false-negative rates;
- comprehensive CDC/RDC, reset, X-safety, or silicon signoff;
- expert consensus; or
- claims about unimplemented properties.
Repository publication sequence
Before labels return, the repository may contain this protocol, rule and manifest definitions, controlled witnesses, tests, limitations, and the packet digest. The packet, candidate identities, mappings, and automated candidate outcomes remain excluded.
After first-pass labels are locked, the repository may add the review scope, reviewer qualification or agreed anonymous description, disclosure, original judgments, disagreements, unresolved cases, repository-risk feedback, and any resulting rule or claim-boundary changes. Candidate RTL is published only after provenance, authorization, licensing, privacy, and blinding review.