Can a database of known crystal structures tell whether a found alloy is new?

E147 — Can a real database settle the novelty question? Probably not, and here is why.

E136 flagged that scripts/search/novelty.py cannot support a discovery claim: its reference list holds only equiatomic alloys, so any non-equiatomic composition reads novel by construction. The queue says to check against a real database instead. One is already local — data/offlattice_hull12.json, 671 phases pulled from Materials Project, JARVIS, NOMAD and Alexandria (E120: 136 prototypes plus 193 alexandria, 167 mc3d, 132 mc3d-psol, 43 jarvis).

But there is a category error waiting here and it should be named before the measurement, not after. That hull is a database of compounds — ordered stoichiometric phases with definite structures. The finds are solid solutions: disordered arrangements on a bcc lattice at arbitrary composition. A solid solution is not the kind of object that appears in a structure database, so absence from the hull is not evidence of novelty. It is evidence of the two being different categories.

Predicted:

1. No in-class find matches any hull phase within 0.15, and the nearest distances are large — above 0.3 for most.
2. The nearest hull phases are binary or ternary compounds, not five-element ones, because the databases are thin on high-order random solid solutions (E14 counted zero MPtrj materials with four or more of these elements).
3. So the hull cannot answer the question. Confirming 1 and 2 establishes that this database is the wrong instrument, not that the finds are new.

Falsified if a find sits within 0.15 of a hull phase, which would mean the databases do carry comparable objects and the hull can serve as a novelty reference after all — a better outcome, and one that would give the project a real check it currently lacks.

E147 result: all three predictions confirmed, and the check found something larger than the novelty question it was asked about.

671 hull phases, by element count:
  1 element   168
  2 elements  503
  3 or more     0

nearest hull phase to each in-class find:
  Mo0.522 W0.151 Ta0.135 Nb0.112 Ti0.080   d = 0.302   WMo2-C15
  W0.412 Mo0.238 Ta0.210 Nb0.077 Ti0.063   d = 0.358   MoW2-C15
  Mo0.589 Ta0.134 Ti0.098 Nb0.092 W0.079   d = 0.263   TaMo2-C15

Predictions 1 and 2 confirmed: nothing within 0.15, nearest 0.263, and every nearest neighbour is a binary Laves prototype. Prediction 3 confirmed: the hull cannot answer the novelty question. It is a database of ordered stoichiometric compounds and the finds are disordered solid solutions — different categories, and absence from one says nothing about the other. The novelty question remains open and needs a literature check, not a structure database.

The larger finding is in the first table. The off-lattice hull — the thing every driving force in this project is measured against — contains no phase with more than two elements. Not one, out of 671.

A convex hull over unaries and binaries can still bound a five-element composition correctly, because the bound comes from tie-planes between phases rather than from a phase at that composition. But it is only correct if no ternary or higher compound is more stable than the best binary mixture. Where one exists and is absent, the hull sits too high, the driving force comes out too negative, and the composition looks more stable than it is.

That is the same failure mode as E54/E56, where C15 Laves phases missing from an earlier hull put seven compositions 98 to 185 meV/atom wrong in the optimistic direction. The current hull has the same blind spot for every ternary and higher intermetallic in this element set — and the in-class finds are all five-element, so they sit squarely in the region it cannot see.

This does not withdraw the in-class finds, because no specific missing competitor has been identified. It does mean their driving forces carry an unquantified one-signed risk, and it should be recorded as a rung-0 blind spot in spec.py, which currently says only that bcc superstructures are excluded.