Proctor reference

The loss, not the whole training system

proctor.py returns cross-entropy, the sequence penalty, and their sum. The weight is fixed at one. Inputs are finite logits and aligned targets for complete, equal-length, unpadded sequences. Computation is float32; there is no target shifting or ignored-position contract.

The original archive and all 22 files are preserved unchanged. The example and tests run on CPU with Python and PyTorch. The package does not include the original corpus, tokenizer, trained weights, or full trainer.

The supplied explainer and JSON records focus on likelihood and concentration measurements. The project page's behavioral account also draws on the original July articles, which describe doom-loop suppression and exploratory capability observations.