ml/__init__.py

"""echo-1 M1 — Stage 0 substrate experiment.

A tiny masked-diffusion language model vs a tiny autoregressive (FIM) baseline,
trained from scratch on pure-Lua infilling, verified by execution. The question
(RQ1): at iso wall-clock latency, does the diffusion substrate's parallel,
bidirectional refinement match or beat left-to-right AR on globally-structured
generation?
"""