rOUT router checkpoints

rOUT is a learned depth router for frozen tabular in-context outlier detectors: given an unlabeled dataset, it reads the backbone's per-layer states of a row subsample and decides after how many layers to stop. These are the routers reported in the paper, three training seeds for each of three backbones.

file	backbone
outformer_s0.pt, outformer_s1.pt, outformer_s2.pt	OutFormer (10 layers)
iclad_s0.pt, iclad_s1.pt, iclad_s2.pt	ICLAD (12 layers)
tactic_s0.pt, tactic_s1.pt, tactic_s2.pt	TACTIC (12 layers)

Each file is a PyTorch dict with the router weights (state), its configuration (cfg), the feature standardization (feat_mu, feat_sd), the feature set, the per-layer cost lam_depth, the stopping threshold chosen on the synthetic validation split for this seed (tau), and the threshold of the three-seed logit ensemble (tau_ensemble).

Usage: put the files into checkpoints/ of the code repository, build the benchmark venues and run scripts/evaluate.sh <backbone> --tau stored (see its README). The backbone weights are not included here; they come from the original releases of OutFormer, ICLAD and TACTIC.