AtmosIQ — Finest Models (v3 final)
Checkpoint: atmosiq_v3_lp_final.pt (289,740 params) — SSL encoder (frozen) + head-only
supervised stage (TP-7 val_mse selection). Encoder: encoder_v3.pt.
M2 spatial-held-out R² (h0, ungauged cells, vs M1 XGBoost floor): pm2_5 +0.303 (xgb 0.30, BEAT) | pm10 +0.363 (0.35, BEAT) | o3 +0.190 (0.13, BEAT) co +0.083 (0.06, BEAT) | nh3 +0.118 (0.06, BEAT) | so2 -0.008 (xgb -0.05) | no2 +0.029 (0.07, below)
Component evidence (not a per-node MLP): graph_gain pm10 +0.041 / o3 +0.085; alpha_self 0.45/0.63; ORPHAN eval (test node's own satellite+static masked, neighbors only): pm2_5 +0.113, pm10 +0.211 — the ungauged-cell claim, demonstrated.
Protocol: TP-1 full pruned-graph training (no node sampling), structural spatial holdout, seed-pinned (digit-level reproducible). λ_phys=0 (covariance proxy refuted by EXP-C, see logs/eval_lp_nophys.txt). Known limits: no2 weakest; σ/ν calibration pending U-1.
Reproduce: scripts/ + tensors/ repo folder + norm.npz + refs.json →
train_v3.py --stage all then evaluate_v3.py.