Layerwise distribution of suffix-sweep effective receptive field. Each layer contains 100 randomly sampled components (seed 0); each component is evaluated at the top-5 rank threshold and summarized by the mean exact-or-bracketed first-recovery length over all dominant-tail occurrences. Suffix lengths 1 through 10 are exact; later lengths use progressively coarser geometric steps and geometric-midpoint estimates. An occurrence that never reaches the threshold is assigned its full available context length. Bars extrapolate sampled fractions to the model's full ICA basis. Panels: gpt2 (GPT-2 small), gemma2 (Gemma 2 2B), qwen9b (Qwen 3.5 9B Base).
