DVIDIA · Verified learning cache preview v0.1

This local preview adds dvidia-train cache for repeated learning-data preparation. Build once from strictly verified camera/action recordings, then load the same derived arrays with independently retained manifest, protocol, split and source pins. Cache hits verify cached bytes; they do not reread or reverify raw recordings.

On one registered 1,000-sample training episode, strict export took 1.636 seconds and verified cache reads took 2.952, 2.831 and 3.214 ms. Numeric features, targets, groups, provenance and the model dataset identity match. This is one warm shared-host preparation benchmark, not a whole-pipeline or robot-control speed claim. Prior measurements and the Windows cleanup correction are retained in the evidence.

A separate numerical preflight checks 4,242 bounded reference/target vectors across all 20 original reference updates. All are within the predeclared 1e-12 native-unit tolerance; 24 negative inputs are rejected. Closed-form projection is not universally bitwise identical. This prepares a future action-representation experiment; no model fitting or native robot attempts were added.

The existing camera-to-action learner remains unqualified at 5/64 strict placements. All 19 frozen runtime files are unchanged. Ordinary video-to-action recovery, reliable manipulation and physical robot transfer remain open.

Install the wheel in a Python 3.12+ environment with the existing pinned NumPy dependency:

python -m pip install ./dvidia_training-0.1.3.dev5-py3-none-any.whl
dvidia-train cache --help

The offline fresh-install checks verify source identity, pinned cache inspect/load, wrong-split refusal and existing robot/footage CLI behavior. Cache construction preserves a 30 GiB disk reserve. No GPU, cloud model or physical-arm command is required by this cache helper.

Usage and benchmark scope · Numeric preflight · Reviewed source PR

Source: cc6746a95669cfd69a49e32a3e6c2e5a42ce2546. Wheel SHA-256: c7335b19d692f3158e8136aa880f9ca2d3d286b2b507703127c88434d7d31c71 (5,826,998 bytes). This is a non-latest engineering prerelease, not a qualified skill package.