TRANSFORMER_BASELINES.md

Transformer baselines

Every promoted LayerCake claim must compare against a matched transformer baseline.

Baseline rules:

• same dataset and split;
• same byte budget;
• matched or disclosed parameter count;
• same hardware;
• same seed count at the declared tier;
• same evaluation script;
• same wall-clock accounting;
• cache state disclosed for generation.

Current implemented baselines:

• TinyByteTransformer for Tier 0/Tier 1 smoke;
• locked 15M BPE transformer artifacts for the north-star CPU/mobile certificate;
• locked receiver BPE transformer artifacts for receiver-frontier verification;
• locked transformer adapter artifacts for Python-domain adaptation comparison.

Tier 2 must use serious matched BPE and byte-transformer baselines with repeated seeds.