Experiments · E236

Do the search's finds change when re-checked with the more accurate energy model?

Not yet. That model was 55 times too slow for the ordering step; the run was stopped with nothing recorded and redone later.

In the log: E223's finds re-walked on the corrected ladder (pre-registered 2026-09-23 11:3x)

stoppedDate 2026-09-23 11:3x, as written in the logrung 4 · DFT3 predictions · 1 result paragraphEXPERIMENTS.md lines 15945–15955, lines 15964–15967
exp E236 diagram
What E236 did and how it came out, drawn from this record and the files it names (book/assets/diagrams/exp/E236.svg).

Pre-registration

  1. (1)
    Finds that show no transition down to 60 K now carry a verdict (p > 0 at rung 1) — at least 10 of the 24.
    no verdict written against it
  2. (2)
    The Mo–Ta-rich finds (Mo₄₂Ta₂₉Ti₂₉, Mo₄₈W₂₉Ta₁₂Nb₈, Mo₅₄Ta₁₆W₁₃Ti₇Hf₅) order > 150 K higher than under v5 (e9 binds B2 Mo–Ta ~30 meV/atom deeper).
    no verdict written against it
  3. (3)
    The top-p pick changes. Caveat stated now: e9 extrapolates worse to unseen compositions (held-48 11.0), so its rung-0 formation energies for unusual compositions carry more model error than v5's; rung 4 (running on the A100 for E223's three picks) is the check. Attribution run (v5 + new grid, isolating the floor fix) queued after.
    no verdict written against it
The pre-registration, as written

E236 — E223's finds re-walked on the corrected ladder (pre-registered 2026-09-23 11:3x)

All 32 walked finds, walk_finds.py --top lattice, E223's environment, with (i) the new rung-1 grid and (ii) rung 0 = the e9 seed ensemble (mean of three), rung 1 on member seed 2 (the best unbiased-set member with a sane held-48). Six shards. Predictions. (1) Finds that show no transition down to 60 K now carry a verdict (p > 0 at rung 1) — at least 10 of the 24. (2) The Mo–Ta-rich finds (Mo₄₂Ta₂₉Ti₂₉, Mo₄₈W₂₉Ta₁₂Nb₈, Mo₅₄Ta₁₆W₁₃Ti₇Hf₅) order > 150 K higher than under v5 (e9 binds B2 Mo–Ta ~30 meV/atom deeper). (3) The top-p pick changes. Caveat stated now: e9 extrapolates worse to unseen compositions (held-48 11.0), so its rung-0 formation energies for unusual compositions carry more model error than v5's; rung 4 (running on the A100 for E223's three picks) is the check. Attribution run (v5 + new grid, isolating the floor fix) queued after.

Results

EXPERIMENTS.md · line 15964

E236 STOPPED (12:2x) — the e9 Hamiltonian is too slow for rung 1. FastECE on a 432-site cell: v5 (embedding 3) 312 full energies/s, e9 5.7 — 55× (cluster functions grow as d² for pairs and d³ for triplets). A rung-1 walk that took ~25 min on v5 would take about a day on e9; after 54 min none of the six shards had finished one find. Nothing was recorded from it.

The full record

This entry is written in 2 separate places in the log, shown here in log order.

EXPERIMENTS.md · lines 15945–15955

E236 — E223's finds re-walked on the corrected ladder (pre-registered 2026-09-23 11:3x)

All 32 walked finds, walk_finds.py --top lattice, E223's environment, with (i) the new rung-1 grid and (ii) rung 0 = the e9 seed ensemble (mean of three), rung 1 on member seed 2 (the best unbiased-set member with a sane held-48). Six shards. Predictions. (1) Finds that show no transition down to 60 K now carry a verdict (p > 0 at rung 1) — at least 10 of the 24. (2) The Mo–Ta-rich finds (Mo₄₂Ta₂₉Ti₂₉, Mo₄₈W₂₉Ta₁₂Nb₈, Mo₅₄Ta₁₆W₁₃Ti₇Hf₅) order > 150 K higher than under v5 (e9 binds B2 Mo–Ta ~30 meV/atom deeper). (3) The top-p pick changes. Caveat stated now: e9 extrapolates worse to unseen compositions (held-48 11.0), so its rung-0 formation energies for unusual compositions carry more model error than v5's; rung 4 (running on the A100 for E223's three picks) is the check. Attribution run (v5 + new grid, isolating the floor fix) queued after.

EXPERIMENTS.md · lines 15964–15967

E236 STOPPED (12:2x) — the e9 Hamiltonian is too slow for rung 1. FastECE on a 432-site cell: v5 (embedding 3) 312 full energies/s, e9 5.7 — 55× (cluster functions grow as d² for pairs and d³ for triplets). A rung-1 walk that took ~25 min on v5 would take about a day on e9; after 54 min none of the six shards had finished one find. Nothing was recorded from it.

Related entries

Built with PRISMWebsite and visualizations made using Claude