Does the machine-learned potential get vacancy energies right inside the MoNbTaW alloy?

E221 — the alloy vacancy anchor, pre-registered (2026-09-21 21:1x; queued after E217 and the E218 finish)

E219's falsified branch: MACE is within 0.06 eV of DFT on pure Mo, so the rung's MoNbTaW value (MACE E_f_eff 3.55 eV against SNAP's 2.48–2.54) is alloy-specific and needs an alloy anchor. scripts/dft/e221_alloy_vacancy_anchor.py: MoNbTaW on the 16 sites of a 2×2×2 bcc cell at Vegard a = 3.2555 Å (four of each species, seed 0), k 7³, 60/720 Ry; the perfect cell and four 15-atom cells — the first site of each species removed — relaxed at fixed cell; E_f(X) = E(15, X removed) − E(16) + μ_X with μ_X the standard pure bcc reference (refs_v3_fit.json); MACE-MPA-0 on the identical cells with its own relaxed pure references; the 1000 K Boltzmann-effective value on both sides (boltzmann_effective, the rung's own estimator). Smoke-tested on the fake pw.x. Predictions: (1) DFT E_f per species 2.4–3.4 eV, Nb/Ta sites below Mo/W (the easy species dominate SNAP's effective 2.5); (2) DFT E_f_eff(1000 K) 2.5–3.0 eV; (3) MACE − DFT within ±0.2 eV on all four species licenses MACE on this basis — trust_kinetics may then be switched on with the measured residual as the correction, and SNAP's 2.5 is the number that is off; > +0.3 eV on any species keeps trust off and the residual becomes the correction the rung needs. About 3 h of pw.x; runs/e221_chain.sh, runs/e221_alloy_vacancy_anchor/results.json.

Li et al. 2024, fcc half imported (2026-09-21 21:3x). scripts/expansion/import_li2024.py --lattice fcc --drop Mn: 9,633 rows of 2–64 atoms (ordered 2–8 and SQS 27/64) in Co, Fe, Cr, Ni, Si, Al, Cu after dropping every Mn-bearing row (25,746; Mn is outside the element universe), 2,228 above 0.5 eV of the hull, 2,537 whose internal relaxation exceeded 0.10 Å (measured by mapping the relaxed cell back onto the initial one and removing the affine strain) and 281 whose volume moved more than 20 %. Their Ef_per_atom is against the most stable unary of each element, and the set holds no pure fcc Fe/Cr/Co/Si to re-reference with; the rows keep that convention (reference: most_stable_unaries) — a cluster expansion absorbs a linear-in-composition offset exactly in its point terms, so a fit on Li alone is unaffected, and joining Li to Waters needs the per-element offsets c_i in E_Li − E_Waters = Σ x_i c_i. scripts/expansion/reference_offsets.py fits them on decorations the two sets share (StructureMatcher on the primitive cells, elements Ni Fe Cr Al Si) and reports the residual — the two codes' settings disagreement — beside the OQMD prior (Fe +161, Cr +394, Si +537, Ni 0, Al 0 meV/atom). Running. Plan: fcc v1 = Waters alone (fitting); v2 = Waters + Li fcc on nine species (Ni Fe Cr Al Ti Si Co Cu — Ti from Waters, Co/Cu from Li), Li rows shifted by the fitted offsets, scored on the same held-out systems plus Co- and Cu-bearing ones; nothing enters the ladder before the pre-registered bars hold.

Reference offsets between the two fcc sets (2026-09-21 21:5x): the sets do not merge on a linear offset. 965 identical decorations (StructureMatcher, Ni Fe Cr Al Si). Fitted E_Li − E_Waters = Σ x_i c_i: Fe +155 meV/atom (OQMD fcc−bcc prior +161 ✓), Si +466 (prior +537), Al +10 (0 ✓), Ni −30 (0), Cr +72 (prior +394 ✗) — and the residual after the offsets is RMS 37 meV/atom, max 208, far beyond a settings difference (ACWF-scale disagreement between PBE codes is 1–3 meV). Cr's implied fcc-minus-ground-state energy of 72 meV is a third of the physical 0.39 eV, and Fe/Cr-rich fcc cells are exactly where spin-polarised relaxations land in different magnetic minima (NM / FM / AFM fcc Cr and Fe differ by 0.1–0.4 eV/atom); the two sets' "same decoration" is not the same state. Decision: the sets are NOT merged. fcc v1 stays Waters-only; a Li-only fcc model (ece_fcc_li_v1, seven species incl. Co and Cu, Li's own convention) is queued behind it; the two are then compared on ORDERING energies — differences between decorations at fixed composition, which no reference enters — on shared compositions, the reference-free cross-check. Open: which magnetic states each set's Cr/Fe cells sit in (Waters carries per-atom moments in relaxed_structures.full.json; Li carries magmom); to be read before either model is believed on Cr- or Fe-rich fcc.

E221 partial — three of four vacancy cells; two predictions already decided. runs/e221_alloy_vacancy_anchor/results.json: DFT E_f (μ = pure bcc) Mo 3.36, Nb 4.31, Ta 4.16 eV; MACE 3.18 / 3.95 / 3.85; W's 15-atom relax never ran: the Colab session dropped at 22:34 ("Session 'forager-qe' not found"), the first job's folder went with the old VM, and the two retries produced empty output (failed 2026-09-23 22:38; attempts parked in vac15_W_relax/colab_outage_2026-09-23/). (1) Falsified without W: Nb and Ta lie above the 2.4–3.4 eV band and above Mo, the reverse of the predicted order. (3) The licence is not granted: MACE − DFT is −0.17 / −0.36 / −0.31 eV, outside ±0.2 on two species, so trust_kinetics stays off. MACE sits below DFT: it overstates vacancy supply, which makes rung 3 err toward reaching a phase change. That is the conservative side for this goal. (2) (E_f_eff at 1000 K) needs W and is not scored. DFT is above SNAP's 2.48–2.54 on every species measured. Whether SNAP's figure uses the same chemical-potential convention has not been checked, so that comparison is not scored. The W cell is re-queued behind E242 on the A100 (runs/e221w_chain.sh).