One row per check, in the same six columns as Tier 1. The difference from Tier 1 is the input: every number here is computed from the latents a Matryoshka SAE learned on this tree, not from constructed statistics. The SAE recovered 20 of 20 true features, which is the ceiling on anything below.
Probe S_res on learned latents
One bar per testable true edge (9 of 9): where the true parent's decoder ranks among the child probe's correlations over the whole dictionary. Rank 0 is the closest. The rule accepts an edge when the rank is inside the top k = 5; an unrelated feature sits at chance in a dictionary of 20 latents.
Children the tree keeps apart
One bar per parent whose children the grammar declares mutually exclusive: how many sampled tokens the SAE's recovered latents co-fire on. The ground truth for every one of these is zero, so the bar is the SAE's own departure from the tree, not a metric's.
Edge recovery
All 9 true edges with the verdict on each. A miss is attributed: child never learned means the SAE has no latent for that endpoint, so the edge was unreachable for any metric.