Coppelius is the 60M-class dense attention transformer inspired by
huggingface.co/ajaxdavis/alpha-60m-base: 16 layers, width 512, 8 heads of
64, FFN 1408, vocabulary 12288, context 1024, RoPE, RMSNorm, tied
embeddings, no matrix biases, and an exact-vocabulary objective.
The physically qualified SM86 realization recorded in docs/WORKLOG.md uses
LayerNorm and a gated GELU feed-forward block because those are the
maintained SM86 owners. The reference family remains explicit here rather
than silently encoding realization-specific padding or substitutions.
The compiler supplied declaration spans and resolved links from this source snapshot. This page does not assert that this file belongs to a checked closure.