Per-head disassemblies

The unified per-head listing for each model — every attention head’s addressing (where it reads) × write (copy/transform) × top content binding × operator role, plus the per-layer MLP read→write neurons. Operator-role tags are hyperlinked to the operator catalog, and load-bearing heads/MLPs carry a ★ discovery-pass causal badge (induction-NLL damage, from the discovered-components sweep). First-order and provisional (single-component reads + the induction idiom).

model	architecture	operator roles linked	causal badges
GPT-2 (small)	GPT-2 / absolute-position	3	15
Gemma-2-2B	RoPE / GQA / RMSNorm	3	21
Llama-3.2-1B	RoPE / GQA / RMSNorm	3	23
Qwen-2.5-1.5B	RoPE / GQA / RMSNorm	3	16

The deeper composition / circuit story is in the circuit catalog and DISASSEMBLY.md. Generated by disassembly_pages.py.