Does retraining the energy model on labels referenced to pure elements fix its DFT misses?

E181 — retrain the eCE on pure-referenced labels (the flywheel's actual first turn)

Same pipeline as v3b in every respect except the label file: _pureref. Composition-wise split, same holdout compositions, same hyperparameters. Name: ece_v4_pureref. Predictions.

1. The four QE cells' DFT/eCE ratios fall from 1.9 / 2.2 / 4.0 / (E180) to within 0.8–1.3 — the eCE re-predicts E172 near +97, E176 near −49, E177 near −71.
2. The composition-wise holdout slope rises from 0.586 to above 0.8; MAE may rise (the label scale is larger) — MAE is not the metric.
3. If the ratios stay above 1.6, the references were not the whole story and the model's own compression is real; that branch measures it against the pure-referenced holdout.

E181 result — predictions 1 (in part) and 2 confirmed; the edge is now the model's. ece_v4_pureref: train/valid/test loss 1.63e-4 / 2.99e-4 / 2.77e-4 eV²/atom² (test RMSE 16.6 meV/atom vs v3b's 14.2 — larger, as predicted). Composition-wise holdout:

model   pred = slope × true + b     r      MAE    true range
v3b     0.586 × true − 9.2        0.759   11.4   [−87, +18]
v4      0.839 × true − 1.4        0.850   15.0   [−112, +53]

Prediction 2 confirmed: slope 0.586 → 0.839. The five QE cells, re-mapped on v4:

cell                          v3b     v4      DFT     DFT/v4
E176  Mo14Nb14Ta13W13        −21.9   −55.5   −49.1    0.89
E177  Mo27Ta18Ti5Hf2V2 s0    −17.9   −78.8   −71.1    0.90
E179  same, seed 1           −17.9   −90.7   −72.5    0.80
E180  Mo27Ta27               −23.6   −96.4   −89.8    0.93
E172  Hf10Mo15Nb2Ta5Ti2V9W1Zr10  +51.2   +12.2   +97.2    7.98

Prediction 1 confirmed for the four refractory-centred cells (0.80–0.93, all inside 0.8–1.3): with labels whose zero is the elements, the eCE predicts DFT formation energies of the search's basin to ~10%. Falsified for E172: at the Hf/Zr-rich edge v4 says +12 where DFT says +97 — the reference fix moved this cell the wrong way (v3b's +51 was closer by accident of the compressed scale). That is the model's own extrapolation failure, now visible because the references no longer hide it — and this cell is also the one whose DFT carries the V-cutoff caveat (9 V atoms at ecutrho 400 < 645). Branch → E172b: the same cell at the settings standard (60/720 Ry, k 4×4×4, refs_v3_std), which separates the two. Prediction: E_form(E172b) within ±10 of +97.2 — the cutoff caveat is small and the +97 stands, leaving the model's edge failure as real and E172 as the first cell that must enter training (the flywheel's proper first turn, on v4). If E172b lands below +70, the cutoff was the caveat and E172's DFT number is withdrawn.

refs_v3_std (60/720 Ry, k-spacing 0.16, 0.96/0.99/1.02): all eight minima bracketed; a₀ = Hf 3.542, Mo 3.167, Nb 3.318, Ta 3.320, Ti 3.265, V 3.007, W 3.185, Zr 3.569 Å — the hcp-formers' bcc constants sit 1–3% below RHEA's lowest-cell values, the refractories within 0.7%. v4 is promoted to the eCE rung for the refractory-centred region; its edge behaviour is the flywheel's first target.