Active-safety mathematics and CPU integration contracts
Five exact Git-tracked source/config/test files preserve the NumPy acquisition/ONS/LURE implementation and two isolated PPO adapter contracts. This is new code relative to the published epoch3 scoring reproduction bundle, not another upload of those models or scripts.
Run python -B -m unittest discover -s tests -v with NumPy and PyTorch on CPU. The math tests load the isolated source file; PPO tests extract two methods by AST without loading a model, DeepSpeed, training runner or network inputs. These checks do NOT prove full trainer dependency closure, policy quality, actual human safety, full three-role resumability or theorem guarantees for PPO. The adapter imports the broader SafeRLHF-V trainer and is provided as exact integration reference, not a standalone end-to-end training system.
The YAML is an algorithm recommendation, not tuned experimental results; it explicitly records zero actual-human observations and not-certified status. No raw queries, human/proxy feedback, sample records, credentials, machine paths, model weights or optimizer state are included. No runner is launched by publication. The historical export script with automatic public-fallback upload behavior is intentionally not bundled or executed.
SOURCE_MANIFEST.json records the exact source commit and file hashes. Read LICENSE and NOTICE; source license does not relicense external model weights or datasets. UPLOAD_MANIFEST.json covers every other release file.