Frozen sample-count benchmark · 8 October 2026

A reproducible CPU benchmark for the current DVIDIA Training source checkout: fit nested sets of 2, 4 and 7 connected native recording groups, install their candidate capsules, and evaluate them on a frozen twelve-layout suite with teacher and matched open-jaw controls.

Nominal qualifications: 8/8, 7/8, 8/8. Actuator-stress qualifications: 3/4 for each count. All 36 open-jaw controls fail; no execution errors were discarded. The seven-group candidate records 0.267 mm mean and 0.483 mm maximum nominal target error. The complete 84-episode benchmark takes 102.727 seconds on one Apple M5 CPU host; scopes are specified in the report.

This is one composition order using aligned simulator labels, privileged state and authored phases/gripper logic. It establishes neither a sufficient human-video count nor hardware qualification. Repeating this seed repeats known layouts.

• Results, protocol, candidate payloads and reproduction guide
• Benchmark tool
• Public source dataset

The attached evidence ZIP retains native traces, qualification/control receipts and replay files; duplicated source media and training intermediates are omitted. SHA256: bea158d27b754ea547693330ddbfacb15b905e32ba8f574341ea1ad8eba192ba (44,732,405 bytes).

This prerelease records an experiment and the current source tool. It does not replace the frozen v0.1.0 pilot or publish a new PyPI package. Install the benchmark from this source revision with the optional [arm] dependencies. The runtime and capsule compatibility contract remain v0.1.0.