| # training_code — single-token-per-step latent-CoT organism | |
| - `latent_threads/single.py` — single-token Markov mask, the T-step in-graph forward with codebook | |
| feedback, readouts (one z_t per step; contrast latent_threads/markov.py). | |
| - `latent_threads/train_single.py` — trainer (answer CE + per-cell feedback CE, teacher-forcing | |
| anneal, mixed chain lengths for length generalisation, separable codebook init). | |
| - `latent_threads/eval_single_report.py` — causal load-bearing battery + length-generalisation curve. | |
| - `latent_threads/eval_single_cellpatch.py` — single-cell `c_i(t):=d` patch + CA-propagation check. | |
| - `latent_threads/plot_single_summary.py` — training curve (readout/state/curriculum vs TF anneal). | |
| - `latent_threads/tasks.py` — the diffuse coupled-CA task. Deps: abstract_cot/masking.py, | |
| model_organisms/envs/base.py. | |
| Retrain: `python -m latent_threads.train_single --config latent_threads/configs/single_k3m6.json --batch-id sg1`. | |