V1 validation evidence

This page records the release-candidate checks from 2026-10-05 and the documentation/package refresh on 2026-10-09, on branch main. It is validation evidence, not a performance claim.

Release surface

• 80 experiment configurations and 16 deployment configurations validated.
• 16 method registrations and 7 environment configuration keys across six task families matched the machine-readable V1 catalogue.
• Soft-robot, co-design, 3DGS, MBD3D, MRMFMBD, and morphology paths were absent from the source and wheel inventories.
• Humanoid push-to-line and MGA GPU remain outside the readiness claim.

Test evidence

The CPU Docker regression produced:

725 passed, 19 skipped, 152 deselected, 0 failed

The deselected cases carry explicit slow, Docker, accelerator, Brax, or Isaac markers. After making D3IL registration lazy, the affected regression added:

6 passed, 2 optional-dependency skips, 0 failed

The validated numerical stack was Python 3.10.20, JAX/JAXlib 0.6.2, NumPy 2.2.6, and SciPy 1.15.3. CI repeats the portable core subset on Python 3.10 and 3.12.

Package evidence

The final release-candidate wheel has:

The wheel and all core dependencies were installed into a new Python 3.12 virtual environment on macOS arm64. pip check, package/version import, JAX CPU backend discovery, optional D3IL plugin import, and the documented planar dry-run all passed from outside the repository checkout.

After the README visual refresh, the wheel was rebuilt and its metadata passed Twine again. Comparing the archives showed changes only in METADATA and RECORD; all runtime/package files are byte-identical to the CPU-tested wheel.

The source and wheel inventory is stored in release/v1_inventory.json. The core dependency license closure is stored in release/v1_dependency_licenses.json.

Static and documentation evidence

• MkDocs built in strict mode.
• 1,017 release Python files parsed with zero syntax errors.
• git diff --check reported no whitespace errors.
• Credential pattern scanning reported zero AWS, GitHub, OpenAI, Slack, or private-key signatures.
• The current source inventory reports three files above 10 MiB: two D3IL camera meshes and the homepage GIF. None is included in the wheel.
• The 20-package core dependency closure reported no detected strong-copyleft license.

Documentation refresh — 2026-10-09

• The homepage now uses 44 original clips in five groups: MDOC, MD-COAS, 2GO, MGA and hardware. Titles, subtitles and rulers were cropped from the inputs; the camera composition adds no text. All five groups and camera views were visually reviewed.
• The 24-second GIF is 896 × 661 at 6 fps (21.57 MiB). The same composition is available as a 1344 × 992, 12 fps MP4 (6.59 MiB). This is presentation playback, not a runtime or latency measurement.
• All 44 input sizes and SHA-256 checksums matched docs/assets/showcase_sources/v2/manifest.json. The exports regenerate from these checked-in inputs without the original research folders or simulators.
• MD-COAS pairs three diffusion views with native animations of 20 candidate trajectories: planar L6/seed 0, planar L10/seed 8 and arm avoidance L1/seed 0. The arm candidates are TCP projections of planned 7-DoF trajectories. Source GIF hashes were verified; every clip occupies exactly one gallery tile.
• The new Python example was run on CPU from the rebuilt wheel, outside the repository checkout: 48 steps, goal error 0.099, with finite-value and goal-arrival assertions passing. pip check and Twine metadata checks passed.
• Planner descriptions were checked against original papers. CFS-MBD is documented as MD-COAS's historical implementation name.
• Environment and constraint catalogues document six task families, seven configuration keys and four numerical constraint-solver interfaces.
• The framework positioning and architecture now connect policy learning, learned trajectory diffusion, model-based generative inference and control. The learning guide documents training, checkpoint reuse and offline reliability calibration without claiming automatic online retraining.
• The roadmap is a concise checklist, and the architecture uses a coordinated learning/task/execution palette. The gallery preserves native footage and a text-free composition.
• The documentation built in strict mode. README links, anchors and Python syntax passed static checks. Training-guide commands were checked against their parsers; no new training runs were performed. Runtime source, task configurations and tests are unchanged by this refresh; the earlier regression remains the runtime evidence.

Manifold positioning and batched constraints — 2026-10-09

• The homepage follows “Generative dynamics on manifold” and shows the complete task-to-execution workflow: task definitions, planners, learning, constraints, the planning workflow, simulation, deployment and evaluation. A rendered PNG accompanies the editable SVG, with Learning explicitly connected to planner priors and offline evaluation feedback.
• The obsolete 2GO corridor and stepping-stone poster images were removed; the unified demonstration gallery is unchanged.
• The batch guide documents candidate, horizon and refinement axes, CBF/CFS and manifold feedback, MGA learned proposals, and DPCC/SafeDiffuser integration boundaries. Its JAX example ran with shape (64, 20, 2) and maximum residual 2.2351741790771484e-08.
• README checks passed for 70 local links, 14 stable section anchors and Python syntax. MkDocs built in strict mode. The rebuilt wheel passed Twine and differs only in metadata from the preceding wheel; runtime files are unchanged.

CI checkout repair — 2026-10-09

• The reported GitHub failure occurred before testing: recursive checkout could not fetch Menagerie commit e5146679f3cfcb327cf759fc9706f5bb0236bd5e from its official remote. This was a local asset adaptation commit.
• The gitlink now pins its official parent a03e87bf13502b0b48ebbf2808928fd96ebf9cf3, verified through a shallow fetch from the official repository. The Go2 MJX changes are preserved as an explicit patch. Setup was checked for first application, repeat application and conflict refusal; the patched XML matches the original local version exactly. The existing local submodule checkout was left untouched.
• A new Python 3.12 core environment ran the workflow's selected tests with coverage: 89 passed, 4 skipped. The selected simulation tests ran in a temporary macOS environment: 20 passed, 1 skipped, including the G1 TorchScript control path. The D3IL startup test skipped for optional dependencies.
• The Ubuntu simulation workflow now installs CPU Torch and urchin explicitly and applies the asset patch after checkout. Test selections remain enabled. These local checks do not claim that the new remote Actions run has completed.

Open gates

• The project now uses MIT. The earlier license inventory recorded the project as UNKNOWN before selection; regenerate that inventory for the release candidate.
• Publication-grade latency, throughput, memory, task-success, safety, and reliability benchmarks still require the versioned hardware protocols and raw artifacts defined in the metrics contract.
• Tagging and publication require a clean-checkout CI run and explicit owner approval.