How do I know the decentralized GPUs I rented actually ran my training job?

Decentralized compute networks let anyone rent GPU capacity to train AI models and receive the resulting weights. The user gets a file, not a proof. Nothing today verifies that those weights came from the training run as specified: an operator could return a pretrained checkpoint, run fewer gradient steps, or swap the dataset entirely. Inference verification, checking a single forward pass, is already hard for ZK proofs; training verification is harder by an order of magnitude because gradient descent is a long sequential process, float16 arithmetic introduces non-determinism that breaks bitwise reproduction, and the training data may be private. Every major DePIN compute network relies on reputation and economic incentives to deter cheating, not cryptographic receipts.