Was the low ordering temperature a coarse-simulation artefact or a real model weakness?

E190b — the discriminator. MoNbTaVW at E188's resolution: 6×6×6 (432 sites), 24 states 1800 → 300 K, 100 uncorrelated samples per state (max 1000), annealed. Prediction: if the sweep was the problem, a peak appears in [550, 950] K; if it still reads ≤ 300 K, reading (i) holds and rung 1 cannot come from v5 as fitted — the within-composition ranking (ρ 0.80 on RHEA's ordered cells) would have to be strengthened with ordered DFT cells before any T_c is trusted, and E188's 500 K for Mo–Ta stands only because Mo–Ta's ordering energy is large enough to survive a weak model. Queued behind E190 (same cores).

E190b result (15:48) — both effects, separated. MoNbTaVW at E188's resolution (432 sites, 24 states 1800 → 300 K, 100 samples/state): T_c = 367 ± 65 K, a clear heat-capacity peak (Cv 3.64 k_B), NN SRO −0.366 at 300 K — where the 128-site / 60-sample sweep had read "still rising at 300 K". Published: 750 (FC17), 742 (WAK23). Neither pre-written branch holds as written: the peak appears (the coarse sweep censored it: resolution effect, real) but at half the published temperature (model effect, real). Reading, with E188 beside it: v5's ordering temperatures come out at ~0.5× the published (Mo–Ta 500 vs 600–1000; MoNbTaVW 367 vs 745; CrTaVW 435 vs 1250 even coarsely) — the within-composition energy scale is weak by about 2×, while the formation energy scale is right to 10 meV/atom. The two are different quantities: formation energy is set by the composition-wise mean the 54-atom random cells pin down; the ordering energy is the spread between decorations, which RHEA's 2–16-atom ordered cells carry with their own rattle-correction noise (12 meV floor against a 29 meV spread). The icet CE's excess (Mo–Ta 1782) and v5's deficit bracket the truth from opposite sides. Consequences: (1) rung 1 stays on hold; (2) any v5 sweep is run at ≥ 432 sites and ≥ 100 samples — the 128-site validation numbers are withdrawn as measurements (kept as the record of a resolution failure); (3) E188's Mo–Ta 500 K is provisional in value but robust in sign: a 2× correction puts it at ~1000 K, the published band is 600–1000 — every estimate except the icet CE's 1782 keeps Mo–Ta's transition inside the window, so the basin's flip from pass to fail stands; (4) branch → task 23: measure the ordering-energy scale directly with DFT — random vs B2-type decorations of one composition in 16-atom cells, minutes each — and feed those cells to v6.

E190b on v5 / v6a / v6b, re-read on the physical axis (2026-09-21 20:2x). The sweeps were 1800 → 300 K nominal = 3600 → 600 K physical. v5 MoNbTaVW: variance peak 367 K nominal = 734 K physical, inside the sweep (dE/dT censored at the floor); α₁ at the floor −0.92. v6a and v6b: both channels censored at the floor, ≤ 600 K physical, α₁ at the floor −0.43 / −0.80. So the E190b/E192 conclusion "the refits moved T_c down" survives the axis change, and the premise behind the whole v6 programme does not: v5's 734 K sits on FC17's 750 / WAK23's 742 (H_mix-inflection and Landau rows; context, not like-for-like) — the "rung 1 is ~0.5× published" that v6a/v6b were built to cure was the sampler's factor of two. The v6 rows (E191/E192 DFT cells) remain valid training data; what they were asked to fix was not broken. v7 (queued) is the first refit to be judged on the physical axis.