The paper and the scripts.

These are the files the evidence ledger cites, served as plain text exactly as they sit in the lab repository. Each script is standalone: Python 3, plus numpy for the four prototypes, and no GPU. Download one and run it with python3 <file>; the ledger lists the arguments behind each published number.

Paper

• Mind-1B.md: the Revision 6 technical specification. It is a design proposal, not a trained model. Links inside it to other lab documents are not published yet.

Constructed-task prototypes

• proto/toy_ternary_lm.py: M0, ternary QAT with a straight-through estimator against fp32 on a character model.
• proto/surprise_gating.py: delta-coded matvec work on a generated redundant stream.
• proto/phase_binding.py: routing by phase synchronization on a seeded synthetic task.
• proto/sleep_cleanup.py: M2, wake/sleep keep-or-prune against a naive retention control.

Planning calculators

• calc/params.py: the ~1.035B parameter planning envelope.
• calc/memory_budget.py: training and inference memory estimates, with every excluded allocation named.

None of these files is evidence of language-model quality. Each card in the ledger states what its script does not establish.