Layer 0 · PCFG toy 4L d_model=448 / pcfg · matryoshka_hook_resid_post_L0 · 1,792 latents in 8 blocks · 1,016,600 tokens over 3400 docs · edge: reverse coverage ≥ 0.5, both endpoints fire ≥ 20
One directory per layer, graded by the same battery as every other source in outputs/. The bar above moves between the layers; nothing in metrics/ differs between them.
Exp 2, axis (b): four formatting densities × three seeds, each its own run — a separate base model and SAE, all trained and graded at layer 2. The density is the fraction of delimiter tokens the grammar emits; the seed row with the bare hash on the compute node is s0.