Chapter 18 — Attention-Head Diagnostics
What a head view covers
Head diagnostics require attention-head structure and the corresponding capture. A declared model without attention does not acquire heads through a setting. A supported model can still lack a head reading on frames where its inputs were not captured or were shed.
Compare effective dimensionality, similarity, motion, and available residual or weight-motion-share readings across the model's actual heads. Use stable head/layer identities and the same measured frame when comparing panels. A head's displayed role label is descriptive; it is not a validated claim that the head implements a named algorithm.
Layer numbering
Read the layer numbering convention with the chart. IRIS™ layer indices are not an additional embedding layer, and model wrappers can expose different module names. Use the spec and the chart's own labels to match a reading to the model rather than inferring identity from screen position.
Coverage and interpretation
An inactive-looking head on the chosen probes is not proof that it is unnecessary on the training distribution. Similar-looking heads are not automatically interchangeable. A concentrated motion share is an observation about that interval, not a statement that the remaining heads are finished.
Check domain and category coverage before interpreting a per-category head view. Keep uncertainty, missing-channel notes, and architecture limits with any exported image. If a head interests you, examine representative probes and independent task behavior before proposing a mechanism.
Attention-head contributions — Figures 6.20–6.24
Where the architecture and captures support them, these figures inspect what individual heads write and compare contributions across pairs of probe groups. Preserve each figure's grouping, frame, coverage, and availability explanation. A measured difference in contribution does not, by itself, establish that a head implements a named behavior or circuit.
