Can we recover how the default twelve-element energy model was built?

EXPERIMENTS.md · lines 7203–7239

E121 — Provenance recovered for the default expansion, and a wrong number in spec.py

E120 found 18 of 26 files in data/ with no producer. Recovering rather than reconstructing, per the rule written into that entry.

Found: the build logs were on disk the whole time, at runs/nickel/ce12.log and ce12f.log, named after the --labels default rather than after what they built, which is why nothing turned them up by name. They record the full invocation, and every parameter checks against the metadata the output files carry:

ce_12element            build_ce_multi.py --extra Ni,Cr,Co,Cu --structures 6000
                        2961 structures, 3081 rows, lasso CV 9.09 meV/atom
ce_12element_focused    ...as above, plus --focus Hf,Mo,Nb,Ta,Ti,V,W,Zr
                        --focus-structures 2000
                        4496 structures, 4616 rows, lasso CV 8.07 meV/atom

The 3081 rows against 2961 structures is a consistency check that passes exactly: the extra 120 are twelve corners at weight 10. Written up in docs/provenance.md.

A number in forager/ladder/spec.py was wrong, and it was mine. Rung 0's error read "cross-validation 8.07 meV/atom on the 12-element fit". 8.07 belongs to ce_12element_focused. The default is ce_12element and its cross-validation is 9.09. Corrected. The file that is meant to be the authoritative record of the ladder carried a 13 per cent understatement of its cheapest rung's error for two hours.

And the better twelve-element fit is not the one in use. ce_12element_focused samples 2000 further structures densely on the refractory eight - the region the generator actually visits - and fits at 8.07 against the default's 9.09. stage_b.build() reads the 9.09 fit. Switching is not a free improvement: every number from E114 onward, including E117's grid and E115's head-to-head, was measured on the 9.09 fit. Recorded as a decision, not taken.

Not recovered, and deliberately not reconstructed: references_equilibrium.npz and references_equilibrium12.npz have no writer anywhere. Their contract is documented from the reader side in docs/provenance.md. The warning there is the important part: the two reference conventions differ by up to 505 meV/atom for vanadium alone (E55), so a producer that yields plausible references rather than these references would silently re-base every energy in this notebook.