# Large-dataset monotonic-depth screen

Screens deep vs shallow `absolute`-residual monotone stacks at each dataset's
full train size, and routes each dataset by the gate (Δ significant *and* ≥ the
practical margin) to a full [size-ladder](loan-size-ladder.md) study or the
standard benchmark. Method: {doc}`protocol` and the
[design spec](https://github.com/davorrunje/mononet/blob/main/docs/superpowers/specs/2026-07-11-large-dataset-screen-design.md).

```{note}
Populated (table + plot) by the GPU session per `benchmarks/RUNBOOK-large-screen.md`.
```