docs/VALIDATION.txt

MICRODUCK BEAK THROW V1 — SOFTWARE VALIDATION
Date: 2026-09-01

POLICY AND ARTIFACT VALIDATION
- beak_throw.onnx: obs[1,61] -> actions[1,14], float tensors.
- alpha_stand.onnx: obs[1,61] -> actions[1,14], float tensors.
- METADATA.json parsed successfully.
- beak_liner_lower_reference.stl: watertight, one component, 1664.0 mm^3.
- beak_liner_upper_reference.stl: watertight, one component, 2438.0 mm^3.
- git diff --check passed for the runtime and handoff bundle.
- microduck-runtime.patch passed `git apply --check --reverse` against the
  modified checkout based on commit 590b986bd8c0d50ae02cb3ea2f59c463b6828168.
- microduck-runtime.patch also passed a forward `git apply --check` against a
  fresh archive of that exact base commit.

RUNTIME TESTS
- duck-control: 60 passed.
- duck-ipc-proto: 49 passed.
- robotctl: 138 passed, 3 ignored visual/manual probes.
- robotd: 98 unit tests passed.
- robotd updater-gate integration: 7 passed.
- robotd-params: 25 passed.
- xtask every_policy_in_the_repo_is_packaged: 1 passed.

The updater-gate integration tests require temporary Unix sockets. Their first
sandboxed invocation was blocked by the sandbox with `Operation not permitted`;
the exact integration target was rerun with socket permission and passed 7/7.

SIMULATOR VALIDATION
- See ACCEPTANCE.txt for the 50-episode randomized throw-to-stand evaluation.
- Dynamic clamped behavior: 50/50 valid releases, landings, forward-carry
  thresholds, lateral thresholds, and recoveries.
- Strict raw-output gate: 0/50 because the policy requested targets beyond an
  anatomical limit before the mandatory runtime clamp.

This validation supports a supervised empty-beak-first experiment. It does not
certify the motion as safe on unmeasured physical hardware.