Merriv v0.1.0a3

Full Changelog: v0.1.0a2...v0.1.0a3

Merriv v0.1.0a2

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.

M2RIV v0.1.0a1

Important

Historical release: this pre-rename M2RIV alpha is preserved for provenance and reproducibility. It is superseded by Merriv v0.1.0a2 and the merriv PyPI package. The old package name and installation instructions below apply only to these archived artifacts.

Reference implementation alpha

This is the first code pre-release of the M2RIV reference implementation for the MCR 0.4 protocol candidate. It is intentionally separate from the evidence-rtx4060-* reproducibility release.

Included

• producer-neutral MCR 0.4 schemas, conformance fixtures, and independent verification
• artifact and numerical diff, paired statistics, release policy, and executed bisect
• recorded CPU quickstart plus the NVIDIA/TensorRT release-evidence vertical
• reusable composite GitHub Action and fail-closed security boundaries

Install

Download the attached wheel and install it locally:

python -m pip install ./m2riv-0.1.0a1-py3-none-any.whl

PyPI publishing is intentionally withheld while the RFC 0008 brand-clearance gate remains open.

Integrity and provenance

The release workflow built the wheel and source distribution from commit abc4530b61b5ac0276d23ef7d90c13942fb8c162, generated the attached SPDX SBOM and SHA256SUMS, and issued GitHub artifact attestations. The Git tag and GitHub-generated tag archives are the canonical source; no separately assembled source ZIP is published.

This is pre-alpha software. MCR 0.4 is a protocol candidate and compatibility may change through the RFC process.

RTX 4060 NVIDIA Evidence Pack (MCR 0.4)

Archived statistical decision

This archive preserves an exact first-party RTX 4060 execution and its original Model Change Report 0.4 output. Its PASS, PASS, BLOCK, BLOCK decisions and build-02-modelopt-int8-scale-065 first-bad result predate Merriv's current matched-binary non-inferiority contract. The policy uses non-zero margins over matched-binary accuracy differences; current Merriv does not apply the continuous sign-randomization assumption to that profile and instead returns ERROR with no current first-bad build.

The retained artifacts, Polygraphy parity, metric values, percentile intervals, runtime observations, provenance, and content identities remain inspectable first-party evidence. Do not cite the archived gate decisions as valid under the current formal profile.

Original execution scope

• Source revision: 073d55b95116e5ef2f420de2e424d5d1c5c29061
• Hardware: NVIDIA GeForce RTX 4060 Laptop GPU
• Driver: 555.97
• TensorRT: 10.4.0
• Polygraphy: 0.53.4
• Paired holdout cases: 629 per build
• Archived decisions: PASS, PASS, BLOCK, BLOCK
• Archived first bad build: build-02-modelopt-int8-scale-065

Trust boundary

This is exact first-party evidence from one Windows/WDDM hardware and software cohort. It is not an independent reproduction, a cross-hardware result, or a performance ranking. The archive retains the artifacts, native Polygraphy output, MCR bundles, and target manifest needed to inspect the observation.

Verification

1. Verify the ZIP with the attached .sha256 file.
2. Extract it outside the source checkout.
3. From a current Merriv source environment, run: merriv mcr verify-target <extracted-target-root>.

The historical M2RIV asset filename is intentionally unchanged so the attached checksum remains an exact reference to the published archive.

ZIP SHA-256: 06A060000AFB40CD9DD6E529B08249863D20A91706030D2B505493572FD21A05