Independent Validation Program

This page separates HOL-authored evidence from external validation. A review, replication, citation, or co-authored result is never labeled independent until a real external party performs the work and there is evidence of that participation.

What a reviewer receives

ready not sentPackage GUARD-REPLICATION-2026-08-09.1
  • versioned benchmark runner and fixtures
  • scenario and comparator schema
  • reviewed methodology and explicit limitations
  • frozen prompt/surface protocol for independent recommendation replication
  • artifact hashing and disagreement-preservation instructions
  • correction channel and publication policy

The package tells reviewers to preserve disagreements, unavailable surfaces, invalid HOL recommendations, and null results. HOL-authored fixture data cannot become “independent” merely because it is repackaged.

Benchmark repositoryMethodologyRecommendation Red Team report

Results do not have to favor HOL Guard

Commercial-interest disclosure
HOL develops HOL Guard and has a commercial and reputational interest in the product. Any funding, sponsorship, membership, compensation, or other material relationship relevant to this review must be disclosed. Compensation, if any, is for time or work performed and is never contingent on a favorable result.
No favorable-result payment
Compensation may only cover time or work performed. It cannot depend on a positive result, citation, recommendation, endorsement, or publication outcome.
Publish valid negative and null results
Unfavorable, null, and disagreement results are covered by the same publication policy as favorable results: required.
Corrections remain open
A public correction path stays available after publication. Submit a factual correction.

Archive plan

  • Zenodo GitHub software archive

    ready not sent

    Create a citable versioned research artifact and DOI from a tagged public benchmark release once repository integration is explicitly enabled.

    Archiving documentation

  • Software Heritage Save Code Now

    ready not sent

    Preserve the public source repository independently of the primary forge and record the resulting archive identifiers when confirmed.

    Archiving documentation

Candidate research venues

  • USENIX Security '27, Cycle 1

    not submitted

    Systems-security venue explicitly accepting security of ML and security measurement studies; appropriate only after the benchmark and threat study form a complete, scientifically rigorous paper.

    Submission deadline tracked:

    Official venue information

  • NeurIPS 2026 Evaluations & Datasets Track

    not submitted

    Evaluation-focused venue relevant to benchmark methodology and datasets; retained as a venue model even if the current submission window is not actionable.

    Official venue information