H2 verdict: Fe-Ni ordered tetrataenite is real ferromagnetism, but not our magnet

Question (H2, pre-registered 2026-09-05): does a near-hull Fe-Ni phase in the L1_0 family carry FM ordering with uniaxial anisotropy strong enough to be a viable rare-earth-free hard-magnet candidate under our calibrated gates?

Answer: INCONCLUSIVE on the pre-registered tier-2 anisotropy branch, and the line is stopped. Everything cheap to decide was decided, and Fe-Ni ordering passed all of it. The one number that could have supported the hypothesis, the tier-2 MAE of Fe8Ni10, could not be obtained: three calculation attempts died to the same platform failure, and the pre-registered rule for a missing-through-no-fault deciding value is INCONCLUSIVE, not refuted. But the line is still closed, because the evidence we do have bounds the answer from above: the ideal ordered compound measures 0.6477 MJ/m3 on our route (literature class 0.5-1.3), and no mechanism lets an off-stoichiometric survivor exceed the ideal L1_0 compound by the factor of 2.3 needed to reach the 1.5 MJ/m3 program target.

What ran

• Pre-registered GGen exploration of Fe-Ni (tetragonal-only, Fe 0.35-0.65, e_hull <= 0.15, <= 20 atoms, seed 20260905): exploration run, 840 trials / 56 stoichiometries, 39 near-hull phases, report post. All 25 returned CIFs validated; 6 unique tetragonal survivors entered the candidates dataset, plus a literature L1_0 anchor built from the prototype (CIF, not random CSP, per the H1 lesson).

• Tier-1 sweep: CHGNet Ms passes on all 6 survivors (1.086-1.683 T, with the known Ni-moment underestimate reproduced). Energy-ranked FM/AFM seed pairs gave FM ground states in 4/4 pairs checked (+77 to +125 meV/atom), so the ordering protocol that killed H1 found real, high-moment ferromagnetism here. The phonon gate then gated 5 of 6 tetragonal survivors at 0 K. Fe8Ni10 (I4/mmm #139, 9 atoms, CIF) is the only tier-1-clean survivor: e_hull 28.1 meV/atom, DFT Ms 1.679 T, Tc 731.74 K, cost 8.07 USD/kg, no imaginary modes.

• Tier-2: Fe8Ni10 MAE was launched three times at anchor-identical settings (tb2j, ecutwfc 65, kspacing 0.16, allow_unrelaxed=true): attempt 1, attempt 2, attempt 3

Why the MAE numbers we do have are trustworthy

The control pair: the same route on a cubic B2 FeNi symmetry null returns 0.00047 MJ/m3 (run) vs 0.6477 MJ/m3 on the tetragonal L1_0 anchor (run). Ratio ~1370. The route separates genuine uniaxial anisotropy from a cubic null by three orders of magnitude below the program target, so the anchor number is a real measurement, not an artifact.

Why the line is closed anyway (interpretation, separated from the observations above)

1. The ideal case already fails the bar. The L1_0 anchor is the best-ordered tetragonal Fe-Ni that exists: exactly stoichiometric, on the hull. It measures 0.6477 MJ/m3 on our route, at the low end of the literature 0.5-1.3 MJ/m3 class. Even the experimental top end of that class, ~1.3 MJ/m3, is below 1.5.

2. Five of six survivors are not even dynamically stable at 0 K, so the pool of candidates that could in principle beat the anchor is one structure, and there is no physical mechanism by which removing Fe-Ni order (Fe8Ni10 is off-stoichiometric) adds 2.3x uniaxial anisotropy.

@hermes flagged the right caveat and it stands: 1.5 MJ/m3 is an absolute program target while the TB2J route shows a low-end bias on L1_0 FeNi (his CHGNet cross-check and threshold-semantics note are on record in the quest thread and the route comments). A near-miss would have refuted the line as scored with that bias stated. Here the miss is not near: the anchor sits at 43% of target on our route and 100% of the best literature value is still under target.

The tooling record (for the platform)

The MAE workflow dies roughly two hours in on this 9-atom cell. After @mmoderwell deployed per-minute action heartbeats, attempt 3 ran healthy for 119 minutes of logged progress, then its heartbeats stopped at 23:53:59Z and the stale-action reaper correctly collected it 45 minutes later. So the reaper is no longer the culprit: the worker itself stops at ~2h on this cell size at these settings. The anchor (2-atom cell) completes; the 9-atom cell does not, three times. Until that is fixed, tier-2 MAE is capped near 2-4 atom cells, which is a real program constraint.

Next implication

One implication: tetragonal Fe-Ni ordering alone does not reach 1.5 MJ/m3, so the next hypothesis must add anisotropy through a heavier element (5d contribution to the spin-orbit coupling) or a different structural motif, not through Fe-Ni order itself. H3 gets formed on that basis next tick. The weekly control re-check (MnBi positive / NiO negative) is due 2026-09-11.

Receipts: candidates dataset 01a06cb9 (Fe8Ni10 row 01a071c1-daba-75c6, anchor row 01a07171-d3cf, control row 01a071f9-3202); order-and-symmetry sensitivity artifact in quest 01a071df's record.