Six-model tiny-run engineering measurements, not policy certification
A numeric-only preservation subset of a Git-tracked historical engineering summary, not new training and not a restored model. The record reports16training prompts,128PTX examples and2saved final actor RL+PTX updates. Original claims are historical measurements, not independently re-run results in this publication.
Policy quality was NOT evaluated. Full three-role RNG/data-cursor continuation was NOT certified. Forced short variable-EOS rollouts served engineering coverage, not policy improvement. Checkpoint checks covered actor same-batch control logits and sampled optimizer tensors, not full three-role continuation. The experiment used a manually assembled runner and CPU optimizer offload; it is not the complete original entrypoint/environment. It retained a native log-space multiplier update, not an exact additive projected update from the paper.
Only explicitly whitelisted numbers, booleans and a numeric length range are included. No prompt, response, feedback, data split, optimizer checkpoint, model weights, internal paths, API payloads or credentials are published. policy_exported=true and checkpoint-saved flags only record historical observations; they do NOT establish current file availability or a usable backup. PROVENANCE.json keeps the summary and recorded runner/commit identities without carrying raw source records. UPLOAD_MANIFEST.json covers every other file.