[0z5a]

The short native recipe defaults to SGD and 48-token responses, so it cannot establish the intended Adam-based online-RL qualification. This adds a frozen, reward-independent 12-train/8-held-out packet and continuous/split/resume commands for GRPO and RLTT, with explicit optimizer, sampling, response-length and checkpoint settings.

The saved-evidence audit requires mixed-reward groups, nonzero masked advantages and finite nonzero gradients together. Version increments or nonzero gradients alone do not pass. Parameter/Adam-state changes, fresh-worker recovery, held-out equality and convergence remain separate required gates.

Stacked on #17 (67da3243); this incremental diff changes no production runtime. Validation: 136 CPU tests pass, and all 20 prompts match the official tokenizer through the actual VIME Dataset. Raw inputs, the corrected input-audit failure and final receipts are retained with SHA256 manifests. The report includes the comparison table and explicit unmeasured cells.

Actual Ray/CUDA execution, the full framework parser, controller execute path, GPU memory/speed and reward convergence have not run for this profile. Revision flags declare intended inputs; actual weights and worker sources still require verification before execution.

Report: native_reward_qualification.md.