Experiments · E187
Yes. The quantum calculation gave −76.4 meV/atom against the energy model's −78.7; the older model was off by 65.
In the log: the fourth flywheel cell, chosen where the rungs disagree most
book/assets/diagrams/exp/E187.svg).Scan of 400 random compositions through the search's rung 0: 21 finds (p ≥ 0.832). Every
one scores −10.5 to −11.7 meV/atom on the icet CE while v5 puts them at −75 to −90 — the
reward the search pays is flat across its own finds. Cell: Mo₂₄Ti₁₀W₂₀ (the largest
disagreement off the Mo–Ta line; v5 −78.7, icet −11.1, icet T_order 40 K), seed-0
decoration, Vegard a from the same table, the settings standard, references refs_v3_std.
Predictions. 1. E_form(DFT) within 15 meV of −78.7 (v5's demonstrated error on six
cells is 3–10). Falsified beyond 15: v5's first miss off its tested lines, and the cell
enters training. 2. The icet CE is off by > 50 (it has been 45–80 off at every find
checked). Queued behind E173 (cores: two searches already running).
E187 result (20:14) — both predictions confirmed. Mo₂₄Ti₁₀W₂₀ at the standard: E_form(DFT)
= −76.4 meV/atom vs v5 −78.7 (Δ 2.3) — prediction 1 (within 15) confirmed on a chemistry off
every line v5 had been tested on; the icet CE's −11.1 is off by 65 — prediction 2 (> 50)
confirmed, and the flat reward's verdict on this find was wrong by the size of the answer.
Seven distinct compositions now, seven within 10 meV/atom of v5. Stored (10 cells in
data/dft/). E191's sixteen-atom ordering cells start on the freed ranks.
Slip, 20:29: the verify() pass-through of the screen's v5 fields landed in
_novelty_factor (same return f, v tail; its s is an int), so every verification raised
from 18:05 on. Lost: E178 seed 2 and E173c at their first call (~10 min each, relaunched);
E193 predates the edit and stands. Fixed by placing the fields in verify()'s own return
dict; both rung-0 modes smoked end to end with --confirm and modulation-without-5-HT.
Colab DFT timing (22:57) — not usable tonight. L4 runtime: 12 vCPU (6 physical), 52 GB, QE 6.7 from Ubuntu's package installs in ~2 min, uploads work into an existing directory in < 8 MB pieces. The 16-atom Mo–Ta B2 cell on 6 ranks: pw.x aborted on rank 1 (SIGABRT) before the first SCF iteration — a packaging fault, not a physics one; not debugged tonight. Session stopped after ~1 h (~4 units). Route if the local queue needs relief: a conda-forge QE 7.x build on the same runtime (~10 min to install), then re-time. Until then every DFT cell runs locally on 6 ranks inside the core and disk budgets.
This entry is written in 2 separate places in the log, shown here in log order.
Scan of 400 random compositions through the search's rung 0: 21 finds (p ≥ 0.832). Every
one scores −10.5 to −11.7 meV/atom on the icet CE while v5 puts them at −75 to −90 — the
reward the search pays is flat across its own finds. Cell: Mo₂₄Ti₁₀W₂₀ (the largest
disagreement off the Mo–Ta line; v5 −78.7, icet −11.1, icet T_order 40 K), seed-0
decoration, Vegard a from the same table, the settings standard, references refs_v3_std.
Predictions. 1. E_form(DFT) within 15 meV of −78.7 (v5's demonstrated error on six
cells is 3–10). Falsified beyond 15: v5's first miss off its tested lines, and the cell
enters training. 2. The icet CE is off by > 50 (it has been 45–80 off at every find
checked). Queued behind E173 (cores: two searches already running).
E187 result (20:14) — both predictions confirmed. Mo₂₄Ti₁₀W₂₀ at the standard: E_form(DFT)
= −76.4 meV/atom vs v5 −78.7 (Δ 2.3) — prediction 1 (within 15) confirmed on a chemistry off
every line v5 had been tested on; the icet CE's −11.1 is off by 65 — prediction 2 (> 50)
confirmed, and the flat reward's verdict on this find was wrong by the size of the answer.
Seven distinct compositions now, seven within 10 meV/atom of v5. Stored (10 cells in
data/dft/). E191's sixteen-atom ordering cells start on the freed ranks.
Slip, 20:29: the verify() pass-through of the screen's v5 fields landed in
_novelty_factor (same return f, v tail; its s is an int), so every verification raised
from 18:05 on. Lost: E178 seed 2 and E173c at their first call (~10 min each, relaunched);
E193 predates the edit and stands. Fixed by placing the fields in verify()'s own return
dict; both rung-0 modes smoked end to end with --confirm and modulation-without-5-HT.
Colab DFT timing (22:57) — not usable tonight. L4 runtime: 12 vCPU (6 physical), 52 GB, QE 6.7 from Ubuntu's package installs in ~2 min, uploads work into an existing directory in < 8 MB pieces. The 16-atom Mo–Ta B2 cell on 6 ranks: pw.x aborted on rank 1 (SIGABRT) before the first SCF iteration — a packaging fault, not a physics one; not debugged tonight. Session stopped after ~1 h (~4 units). Route if the local queue needs relief: a conda-forge QE 7.x build on the same runtime (~10 min to install), then re-time. Until then every DFT cell runs locally on 6 ranks inside the core and disk budgets.
Built with PRISMWebsite and visualizations made using Claude