Do proposals from a second, independent generator run also pass every check?

EXPERIMENTS.md · lines 3642–3686

E67 — Eight of eight, from two independent seed sets, and the best is a binary found twice

E62 verified four compositions the fly proposed, all from seeds 0 to 7. Whether that was the generator or that particular run's luck was untested. Four more were taken up the full ladder from seeds 100 to 107, which share nothing with the first set.

Predicted before running: they would be Mo/Nb/Ta/W-rich, or the convergence is not reproducible; and MACE would come in 14 to 42 meV/atom more stable than the screen, the bias measured across nine earlier compositions.

All four are 100 per cent Mo, Nb, Ta and W - the same corner the first set occupies, and the corner E59 found holds 0.41 per cent of the simplex. Prediction one holds outright.

Prediction two nearly holds. The gaps were 37, 16, 24 and 43 meV/atom; three inside the band and one a meV outside it. Useful as a rule of thumb, not clean.

The binary was found twice, independently. Seed 1 produced Ta.53 Mo.47 and seed 101 produced Mo.51 Ta.49 - the same two-element alloy, from runs sharing no random state. Both pass the same way, with an ordering transition above the service window rather than inside it, which means they enter service already ordered and never change. That satisfies the requirement as written and makes them B2 intermetallics rather than solid solutions, with whatever that implies for ductility. The ambiguity is the operator's to resolve.

One candidate passes by the other mechanism entirely. Ta.49 Mo.20 Nb.16 W.15 orders at 453 K, the lowest of any composition examined, which puts its transition deep in the range where nothing moves. It is a genuine solid solution held by kinetics, not an intermetallic held by being already ordered - and it is the only one of the eight that is.

Eight of eight pass. Every composition the generator has proposed and had checked survives all three rungs, and the two best are roughly twice as stable against decomposition as the alloy the literature cites.

What this does not establish. Every number is MACE and the cluster expansion; no DFT has touched any of it, because none is installable on this machine. The ordering temperatures carry the Monte Carlo scatter only. And the fly matching a hill-climber on search efficiency (E61, p = 0.138) is unchanged by any of this - what is established here is that what it proposes survives checking, not that it proposes it faster.