WAMProbe v0.1.0

WAMProbe v0.1.0 is the first stable release of the counterfactual evaluation toolkit for
World Action Models.

Highlights

• typed, dependency-free CPU evaluation core for action-conditioned state futures
• PointMass, BlockPush, Gripper-Catch, LIBERO-CF-Mini, video/control, and closed-loop studies
• causal consistency, dynamics, ranking, regret, uncertainty, and control-utility profiles
• pinned StarWAM + Wan2.2 integration evidence and explicit unsupported-capability skips
• English and Simplified Chinese READMEs with real report visuals
• tested custom adapter starter, contributor guide, structured result forms, and roadmap

Install

python -m pip install wamprobe==0.1.0
wamprobe demo --contexts 12 --seed 7 --output runs/pointmass-demo

Verification

• Python 3.11, 3.12, and 3.13 CI passed
• 79 tests passed with 88.35% coverage
• wheel and sdist reproduced byte-for-byte from the tagged commit
• archive safety, metadata, committed evidence hashes, and clean offline wheel install passed
• GitHub build-provenance attestations were generated from tag v0.1.0

Artifact SHA256:

• wamprobe-0.1.0-py3-none-any.whl: 8563a8a6e945aeaab71401b72e03a4173e75f01e7830e2fc24d5a9b7c4f661de
• wamprobe-0.1.0.tar.gz: c7aeda26c9747ccfd281ba73928e1e325f3245409fa62bd6d35539bcbef6da80

Reproduction status

Maintainer clean-install checks from public artifacts are recorded in Issue #2. These
owner-run checks are maintainer acceptance and are not claimed as independent external
reproduction.

See CHANGELOG.md for the full release history and release-manifest.json for the audited
artifact/evidence record.

WAMProbe v0.1.0rc1

Changelog

All notable changes to WAMProbe will be documented here.

[0.1.0rc1] - 2026-07-15

This candidate is prepared for reproducibility review and is not yet published to PyPI or
as a final GitHub Release.

Added

• Strict MkDocs documentation, GitHub Pages deployment, repository-local Markdown link
  validation, and Draft 2020-12 validation of canonical public-schema instances.
• Typed alpha WAMAdapter protocol and capability declaration.
• Analytic PointMass-2D paired counterfactual benchmark.
• Oracle, copy-last-frame, wrong-direction, and action-agnostic reference baselines.
• Action Dependence, Counterfactual Direction Accuracy, No-op Stability, state ADE,
  and Top-1 Regret metrics.
• JSON/Markdown reports, schemas, CLI, CI, and contributor documentation.
• Phase 0 evidence map covering 15 reported WAM/VLA failure modes.
• Version-pinned audit selecting StarWAM as the first real adapter target.
• Provider-scoped local model store with pinned revisions, expected paths, sizes, and
  checksums for the first StarWAM integration.
• wamprobe doctor with safe manifest parsing, revision/size/hash validation, and JSON
  output for local model artifacts.
• Reproducible StarWAM environment preflight and a narrow patch that restricts verified
  inference checkpoint loads to PyTorch's weights-only deserializer.
• Checksummed intervention JSONL, corruption-detecting content-addressed prediction cache,
  exact-context comparison, and report-only regeneration commands.
• Contact-aware BlockPush-2D and attachment-aware Gripper-Catch analytic benchmarks.
• Four-family LIBERO-CF-Mini generation with exact snapshot restore, repeat, branch-order,
  provenance, and cache validation.
• A pinned StarWAM 36-prediction matrix and 36 action-chunk simulator executions with
  explicit unsupported-metric skips and retained zero-success results.
• Dependency-free PSNR/global-SSIM diagnostics and a traditional-video/control-value
  counterexample study.
• Minimal score-execute-observe closed-loop replanning with context-block intervals and
  offline-ranking/return association analysis.
• Release artifact audit, candidate workflow, clean-wheel smoke test, evidence manifest,
  and technical-report draft.

Changed

• GitHub workflows now use the current Node 24-compatible major releases of official
  checkout, Python setup, artifact, Pages, provenance, and uv setup actions.