Did the fix to the pure-element reference energies actually reach the model in use?

EXPERIMENTS.md · lines 2890–2931

E55 — The reference fix reached the source and never reached the data

E53 found that every element had been referenced at one shared lattice constant, 3.2935 A, charging the alloy for elastic energy no element pays in its own crystal - 505 meV/atom for V, 486 for Zr, 388 for Hf - and concluded: "References are now each element's own equilibrium." scripts/build_ce.py was indeed corrected, and it wrote data/references_equilibrium.npz on 14 September.

The expansion in use, data/ce_8element.npz, is dated 12 September. It was never rebuilt. Element by element, its stored references match the shared-lattice column exactly and the equilibrium column not at all:

which is E53's own table of what the strain charged, reproduced to the last digit. The fix was written down and the artefact kept the bug. Every mixing energy the pipeline has produced since is still too negative by 226 +/- 43 meV/atom.

Found by accident, and that is the point. It surfaced only because the off-lattice hull (E54) needed absolute energies and the elemental bcc cells were relaxed with MACE, giving an independent measurement of the same eight numbers. Nothing in the pipeline compares its stored references against anything, so nothing could have noticed.

A refit is not needed. Changing a reference shifts every fit target by a function linear in composition, and the singlet terms of a cluster expansion are themselves linear in composition, so the shift is absorbed exactly by the singlets and no pair or higher term moves. CEThermo.correct_references adds it back at load. Measured at equiatomic the correction is 224.8 meV/atom, against E53's predicted 226 +/- 43.

What moves, and what does not. Transition temperatures are untouched, for the same reason the correction is exact: they come from the energy variance at fixed composition, where any reference is a constant and cancels. The survey stands. E54 stands, since it never routes through the expansion's references at all - the solid solution there is relaxed directly with MACE.

What moves is the sign of the answer. The equiatomic mixing energy goes from -152 meV/atom to +73 meV/atom: not a favourable mixture at all, but one that would rather separate or order at zero temperature and is held together by entropy. That is the accepted picture of a refractory high-entropy alloy, and the number the pipeline had been reporting contradicted it.