Are the database's energies for strongly ordered cells right, checked against our own DFT?
Partly. As used, all four sit within 13 meV/atom of DFT; the model, not the data, missed them. One corrected label (Cr–Ti) is 21.9 off.
In the log: are RHEA's deep ordered labels right? (pre-registered 2026-09-22 20:1x)
mixedDate 2026-09-22 20:1x, as written in the logrung 4 · DFT3 predictions · 1 result paragraphEXPERIMENTS.md lines 15826–15842, lines 15929–15935
What E232 did and how it came out, drawn from this record and the files it names (book/assets/diagrams/exp/E232.svg).
Pre-registration
(1)
|QE − corrected| ≤ 10 on
all four.
failson Cr–Ti only (21 …
(2)
On Ta–Ti and Cr–Ti — where the median Ti reference moves the label by 21 and
28 — QE sits nearer the corrected label than the one v5 saw.
holdsfor Ta–Ti, fails for …
(3)
Falsifier: any cell > 20 off
means the data carries part of the ordered-state error (for Cr–Ta that would reopen E222's
reading). scripts/dft/e232_rhea_ordered_audit.py (fake-pw.x smoke test in
tests/test_dft_scripts_smoke.py); chain runs/e232_chain.sh after E230, ≈ 1 h of pw.x.
no verdict written against it
The pre-registration, as written
E232 — are RHEA's deep ordered labels right? (pre-registered 2026-09-22 20:1x)
Only one RHEA ordered label has ever met this project's QE (B2 Mo–Ta, −183.9 vs −184.5). v5
misses its own deepest ordered 16-atom rows by +56.8 (Cr–Ta, B2), +18.9 (Ta–W), +15.3 (Ta–Ti),
−11.3 (Cr–Ti) meV/atom, and a ridge regression of its training residuals on excess neighbour
pairs gives Mo–Ta +11.7 (NN) / −11.3 (2NN) and Cr–Ta +25 — v5 under-fits order on its own data.
If the labels are right, every ordered-state miss is the model's and the refit programme is a
model programme. Cells: those four RHEA decorations on ideal sites at Vegard a, QE v3
(60/720 Ry, k 7³, non-spin as RHEA), formation against refs_v3_fit. Labels frozen before any
DFT (runs/e232_labels.json, E231's code path): as v5 saw them / E228-corrected —
Cr–Ta −37.6 / −42.4, Ta–W −89.1 / −86.5, Ta–Ti +25.5 / +4.8, Cr–Ti +77.0 / +49.5
(B2 Mo–Ta −183.9 / −183.5 against QE −184.5). Predictions. (1) |QE − corrected| ≤ 10 on
all four. (2) On Ta–Ti and Cr–Ti — where the median Ti reference moves the label by 21 and
28 — QE sits nearer the corrected label than the one v5 saw. (3) Falsifier: any cell > 20 off
means the data carries part of the ordered-state error (for Cr–Ta that would reopen E222's
reading). scripts/dft/e232_rhea_ordered_audit.py (fake-pw.x smoke test in
tests/test_dft_scripts_smoke.py); chain runs/e232_chain.sh after E230, ≈ 1 h of pw.x.
Results
EXPERIMENTS.md · line 15929
E232 RESULT. QE vs RHEA's deepest ordered cells, (label v5 saw / corrected): Cr–Ta −35.2 vs
−37.6 / −42.4; Ta–W −85.8 vs −89.1 / −86.5; Ta–Ti +12.7 vs +25.5 / +4.8; Cr–Ti +71.4 vs +77.0 /
+49.5. (1) fails on Cr–Ti only (21.9 off the corrected label); (2) holds for Ta–Ti, fails for
Cr–Ti; (3) fires on Cr–Ti. RHEA's ordered labels are within 13 meV/atom of QE on all four — v5
missed the same rows by up to 57, so the ordered-state error was the model's (E229 agrees). The
median Ti reference is not settled: it helps Ta–Ti and hurts Cr–Ti; three strained Ti frames do
not pin it.
The full record
This entry is written in 2 separate places in the log, shown here in log order.
EXPERIMENTS.md · lines 15826–15842
E232 — are RHEA's deep ordered labels right? (pre-registered 2026-09-22 20:1x)
Only one RHEA ordered label has ever met this project's QE (B2 Mo–Ta, −183.9 vs −184.5). v5
misses its own deepest ordered 16-atom rows by +56.8 (Cr–Ta, B2), +18.9 (Ta–W), +15.3 (Ta–Ti),
−11.3 (Cr–Ti) meV/atom, and a ridge regression of its training residuals on excess neighbour
pairs gives Mo–Ta +11.7 (NN) / −11.3 (2NN) and Cr–Ta +25 — v5 under-fits order on its own data.
If the labels are right, every ordered-state miss is the model's and the refit programme is a
model programme. Cells: those four RHEA decorations on ideal sites at Vegard a, QE v3
(60/720 Ry, k 7³, non-spin as RHEA), formation against refs_v3_fit. Labels frozen before any
DFT (runs/e232_labels.json, E231's code path): as v5 saw them / E228-corrected —
Cr–Ta −37.6 / −42.4, Ta–W −89.1 / −86.5, Ta–Ti +25.5 / +4.8, Cr–Ti +77.0 / +49.5
(B2 Mo–Ta −183.9 / −183.5 against QE −184.5). Predictions. (1) |QE − corrected| ≤ 10 on
all four. (2) On Ta–Ti and Cr–Ti — where the median Ti reference moves the label by 21 and
28 — QE sits nearer the corrected label than the one v5 saw. (3) Falsifier: any cell > 20 off
means the data carries part of the ordered-state error (for Cr–Ta that would reopen E222's
reading). scripts/dft/e232_rhea_ordered_audit.py (fake-pw.x smoke test in
tests/test_dft_scripts_smoke.py); chain runs/e232_chain.sh after E230, ≈ 1 h of pw.x.
EXPERIMENTS.md · lines 15929–15935
E232 RESULT. QE vs RHEA's deepest ordered cells, (label v5 saw / corrected): Cr–Ta −35.2 vs
−37.6 / −42.4; Ta–W −85.8 vs −89.1 / −86.5; Ta–Ti +12.7 vs +25.5 / +4.8; Cr–Ti +71.4 vs +77.0 /
+49.5. (1) fails on Cr–Ti only (21.9 off the corrected label); (2) holds for Ta–Ti, fails for
Cr–Ti; (3) fires on Cr–Ti. RHEA's ordered labels are within 13 meV/atom of QE on all four — v5
missed the same rows by up to 57, so the ordered-state error was the model's (E229 agrees). The
median Ti reference is not settled: it helps Ta–Ti and hurts Cr–Ti; three strained Ti frames do
not pin it.
Related entries
E231 — the E228 relabel, one variable: v5's rows, v5's recipe, corrected labels (pre-registered…
E228 — Is the correction under every RHEA label too soft? (pre-registered 2026-09-22 19:1x)