Can a few dozen of our own DFT cells pin down each element's reference energy?

E239 — pin the label frame's pure-element references with QE anchor cells (pre-registered 2026-09-23 16:3x)

Problem. The eCE's formation-energy labels subtract one RHEA pure-bcc reference per element. Ti, Cr and Hf are poorly pinned: Ti has 3 pure frames, Hf's 12 corrected frames spread 34 meV/atom, and the two defensible estimators (median, minimum) differ by 33.5 / 28.1 / 22.0 meV per unit fraction of Ti / Cr / Hf. A MACE-free route (DFT-force harmonic correction + E(V) fit on the pure frames) was tried and fails: the Ti and Hf frames span a = 3.0–3.6 Å and are heavily rattled; even Mo scatters 28 meV that way. The error is composition-linear, so it cancels in every energy difference at fixed composition (ordering, rung 1) and moves only absolute formation energies (rung 0's drive and hull screen). Fix: measure it. ~48 of RHEA's own ordered cells, recomputed in QE at the v3 standard (60/720 Ry, k 0.16 Å⁻¹, MV 0.02, non-spin as RHEA — E232's settings), each on ideal sites at the label's relaxed volume; formation energy against refs_v3_fit. Then E_form(QE) − E_form(label) = Σᵢ xᵢ δμᵢ + ε, least squares, bootstrap intervals. Selection, fixed before any DFT: RHEA bcc_alloys_ordered rows in the final label set that are trusted, have |correction| ≤ 30 meV/atom and ≤ 16 atoms; stratified so every element appears in ≥ 8 cells and every Ti/Cr/Hf pairing with the other eight elements in ≥ 1; seed-fixed. None of these cells is in the eleven QE cells or E223's three, which stay the independent test. Predictions. (1) δμ for Mo, Nb, Ta, V, W is within ±5 meV (well-pinned references). (2) The per-cell residual sd after the fit is ≤ 10 meV/atom (E232: four cells within 13). (3) At least one of δμ_Ti, δμ_Cr, δμ_Hf differs from zero by > 2σ, and the fitted values land nearer the MEDIAN references than the minimum (my prior, not a certainty — the E238 early read points the other way). Decision: every δμ that is > 2σ from zero is applied to the labels as a composition-linear correction; rung 0 is refitted on the corrected labels and scored on the unchanged independent cells; rung 1 needs no refit for a pure reference shift (checked by the E238 diagnostic's ordering rows). Script written by Fable, scripts/dft/e239_anchor_references.py; runs on the A100.

E239 interrupted 19:56–20:31 (20:3x record). A rewrite of the shared colab_pwx.py drop-in dropped its executable bit; E239's launch of cell i16555 failed with PermissionError and the chain stopped after 10 of 48 cells. Restarted at 20:31 (resume-safe: the 10 finished cells were read from disk, the empty cell relaunched); first-run log kept as runs/e239_anchor_references.part1.log. A test now asserts the drop-in stays executable. The spin-ladder known-answer run's local driver died in the same window; its A100 stage job kept running and was re-adopted (runs/scf_spin_test/adopt_orphan.py).