Skip to content

Merriv v0.1.0a2

Pre-release
Pre-release

Choose a tag to compare

View all tags
@niansia niansia released this 30 Aug 10:49
· 13 commits to main since this release
v0.1.0a2
cff5951

Known statistical limitation

0.1.0a2 predates the matched-binary non-inferiority correction now present on main. Do not use this release for formal Holm-Bonferroni gating of matched-binary metrics with a non-zero margin: its centered sign-randomization test assumes sign exchangeability that the shifted {-1, 0, +1} support generally does not satisfy. Current development code fails that profile closed as ERROR until a dedicated matched-binary non-inferiority method is adopted.

Use 0.1.0a2 to evaluate installation, evidence formats, conformance, and workflow mechanics—not as production binary non-inferiority decision evidence. Model Change Report 0.4 wire identity is unchanged.

Merriv reference implementation alpha

This is the first pre-release under the unified Merriv name. The Python distribution, import namespace, CLI, repository, and release artifacts are all merriv.

Install from PyPI

uvx --python 3.13 --from merriv==0.1.0a2 merriv --help

Included

  • policy-wide Holm-Bonferroni correction, target power, MDE, and fail-closed INSUFFICIENT_POWER
  • source-anchored replay of llama.cpp issue #22544
  • Model Change Report 0.4 conformance and Python/Node/Rust content identity
  • in-toto predicate/Statement and OCI 1.1 referrer-layout prototypes
  • Polygraphy importer and retained native evidence boundary

Integrity and provenance

The protected release workflow built the wheel and source distribution from commit cff5951cdd25183cf3da9fe3cfa4ddd95b58a028, generated the attached SPDX SBOM and SHA256SUMS, issued GitHub artifact attestations, and published to PyPI with Trusted Publishing. The Git tag and GitHub-generated tag archives remain the canonical source.

This is pre-alpha software. Model Change Report 0.4 is an external-review baseline, not a claimed standard.

Assets 6
Loading