Experiments · E234
Not yet. The cells were labelled (2,031 of 2,059), but the fit crashed on a data bug; they went into later fits instead.
In the log: the ordered cells the labeller dropped (pre-registered 2026-09-22 20:3x)
book/assets/diagrams/exp/E234.svg).RHEA holds 4,750 bcc_alloys_ordered frames; 2,691 are cubic 2m³ cells and were labelled,
2,059 (3–5-atom enumerated orderings, the data closest to what the eCE paper trained on) were
skipped because ideal_frac only maps cubic cells. Generalised today: any bcc supercell in the
standard orientation maps exactly by rounding to (a/2)(n₁,n₂,n₃), n all even or all odd, verified
(distinct sites, ≤ 0.7 Å); the energy side scales the frame's own cell isotropically (identical
for cubic frames; 9 tests). 2,031 of 2,059 map; 28 do not and are counted. Their atoms are on
the ideal sites (median and 90th-percentile displacement 0.000 Å, max 0.55), so the correction
is pure strain — the cleanest ordered labels in the set. Labelled at the true minimum into
runs/rhea_labels_noncubic.jsonl (running).
Fit (after E233): E231's refined v5 rows + these cells (|a_in − a_relaxed| ≤ 0.10), pass 2
with the refined median references, v5's recipe, seeds 0–2 → ece_v5o_seed*, one variable (data)
against the E231 ensemble. Mixing them with unrefined v5 rows is forbidden: those sit ≈ 17
meV/atom high and the new cells do not, which would bias ordered against random by exactly that.
Predictions. (1) Ensemble MAE on the eleven unbiased cells from E231's value to ≤ 14, the
four Mo–Ta cells' mean to < +15, B2 Mo–Ta to < +10. (2) held-48 (refined labels) within
1.0 of E231's. (3) Falsifier: eleven-cell MAE within ±3 of E231's → ordered data alone does
not move the eCE, and the pair range (E233) or the embedding (E229) must. Chain runs/e234_chain.sh.
No result paragraph for this entry was found in the log.
RHEA holds 4,750 bcc_alloys_ordered frames; 2,691 are cubic 2m³ cells and were labelled,
2,059 (3–5-atom enumerated orderings, the data closest to what the eCE paper trained on) were
skipped because ideal_frac only maps cubic cells. Generalised today: any bcc supercell in the
standard orientation maps exactly by rounding to (a/2)(n₁,n₂,n₃), n all even or all odd, verified
(distinct sites, ≤ 0.7 Å); the energy side scales the frame's own cell isotropically (identical
for cubic frames; 9 tests). 2,031 of 2,059 map; 28 do not and are counted. Their atoms are on
the ideal sites (median and 90th-percentile displacement 0.000 Å, max 0.55), so the correction
is pure strain — the cleanest ordered labels in the set. Labelled at the true minimum into
runs/rhea_labels_noncubic.jsonl (running).
Fit (after E233): E231's refined v5 rows + these cells (|a_in − a_relaxed| ≤ 0.10), pass 2
with the refined median references, v5's recipe, seeds 0–2 → ece_v5o_seed*, one variable (data)
against the E231 ensemble. Mixing them with unrefined v5 rows is forbidden: those sit ≈ 17
meV/atom high and the new cells do not, which would bias ordered against random by exactly that.
Predictions. (1) Ensemble MAE on the eleven unbiased cells from E231's value to ≤ 14, the
four Mo–Ta cells' mean to < +15, B2 Mo–Ta to < +10. (2) held-48 (refined labels) within
1.0 of E231's. (3) Falsifier: eleven-cell MAE within ±3 of E231's → ordered data alone does
not move the eCE, and the pair range (E233) or the embedding (E229) must. Chain runs/e234_chain.sh.
Built with PRISMWebsite and visualizations made using Claude