Do independent models agree on one surviving find worth testing with DFT?

EXPERIMENTS.md · lines 16241–16252

E241 — pick the DFT survivor by agreement across independent models (pre-registered 2026-09-23 20:2x, before E240 finished)

After Blalock … Romero (bioRxiv 2026.09.15.751856): optimise the compromise between independent models, never one learned model that extrapolates. Eligible: E240 finds whose ladder p ≥ 0.5 under E237's e6 ensemble (rung 1 finds no ordering in 90–1000 K; rung 2 keeps it a solid solution). Three independent votes: (a) MACE rung 2: drive_cold_rung2 < 0 (the solid solution below the cheapest off-lattice mixture); (b) the empirical rules from tabulated element data: δ < 6.6 % and VEC < 6.87 (forager.physics.descriptors); (c) v5 (E223's walk, a different model and labels): no transition found inside 90–1000 K (NaN, or the old "none above 100 K"). Survivor: most votes, then highest p, then most negative drive_cold_rung2; if nothing earns all three votes, the best is taken and the missing vote stated. The survivor goes to an E222-style DFT test on the A100 (ordered vs random cells at its composition; settings pre-registered at launch). scripts/validate/e241_agreement.py.

EXPERIMENTS.md · lines 16287–16289

E241 RESULT (pre-registered rule, on the re-scored walk): 11 eligible, 10 with all three votes; survivor Mo₅₁Ti₃₈W₃Ta₃ (p 0.771; MACE drive −110.0 meV/atom; δ 3.5 %, VEC 5.13; v5 found no transition in the window).