Status: proposed / publication-ready library memo\nPublication date: 2026-06-20\nAudience: materials labs, MLIP builders, and observers evaluating the Lupine evidence trail.
\nUniversal machine-learning interatomic potentials and generative materials models have become good enough to change the substrate of materials discovery. The bottleneck is no longer only “can we generate or relax more candidates?” It is “which prediction is reliable enough to spend DFT, synthesis, or lab time on?”
\nLupine’s next public lane should frame itself as the correction and trust layer for that substrate. The system should not claim to have discovered a new material yet. It should publish evidence-carrying canaries: sealed candidate sets, paired baseline-vs-correction comparisons, property-specific guards, refusal logs, and kill criteria.
\nRecent public work points in the same direction:
\nThe Lupine article thesis should be simple: foundation models make screening cheap enough that correction, calibration, and evidence become the scarce layer.
\nA single MLIP accuracy number is not a discovery certificate. Each material candidate should carry separate trust status for the property being used:
\n| Trust contract | \nWhat it certifies | \nWhat can break it | \n
|---|---|---|
| Relaxation trust | \nrelaxed geometry is physically plausible | \nhidden phase change, unstable tensor, force explosion | \n
| Energy-ranking trust | \ncandidates are ordered correctly enough for screening | \nfunctional bias, stoichiometry bias, support mismatch | \n
| Force/MD trust | \nfinite-temperature trajectory is usable | \ndrift, rare-event failures, bad local curvature | \n
| Phonon trust | \ndynamical stability and second derivatives are credible | \nimaginary modes, displacement sensitivity, Hessian collapse | \n
| Migration-barrier trust | \nNEB / hop barriers are decision-relevant | \nwrong transition state, barrier ranking inversion | \n
| Surface/catalysis trust | \nadsorption and reaction surfaces are credible | \nbulk-training extrapolation, charge/coverage effects | \n
| Pressure/temperature trust | \nmodel survives thermodynamic extrapolation | \nout-of-domain pressure, phase-boundary instability | \n
| Electronic-property trust | \nstructure is suitable for band/optical claims | \nMLIP geometry confidence does not imply electronic accuracy | \n
This is the strongest discovery-facing wedge. The repo already contains a LiFePO4 local canary where a paired correction improves final position RMSE from about 0.131 Å to about 0.042 Å under a force guard. That is not a conductivity discovery, but it is exactly the kind of sealed paired result that can mature into a battery trust lane.
\nNext gate: lock a public solid-ion-conductor reference subset such as Li3YCl6 or Li6PS5Cl, run MACE/CHGNet/ORB/SevenNet baselines and corrections, and score position, force, stress, energy drift, migration proxy, and refusal rate.
\nKill if the correction improves position while worsening force/stress, fails shared-checkpoint pairing, or cannot lock a redistributable reference.
\nThe MPtrj canary is the best bridge beyond the controlled Ni fixture. The support-floor v2 cloud result reports eight paired comparisons, six improvements, two safe holds, and zero regressions. That is a useful evidence-contract result, not a universal model-improvement result.
\nNext gate: replay the row-hybrid v3 policy in Cloud Run with force/stress/elastic rows promoted to first-class gates.
\nKill if the win remains energy-only while the article implies broader model improvement.
\nNi is a credibility anchor. The broad Ni paired-accuracy campaign rejected promotion: 25 measured pairs, zero improvements, ten regressions, and fifteen unchanged. That negative result is useful because the gate caught harm. A local Ni vacancy canary is promising but not yet externally locked.
\nNext gate: lock vacancy formation, local relaxation, and migration references before promoting any defect-transport claim.
\nKill if bulk anchors regress or the local improvement fails cloud/shared-checkpoint replay.
\nFe remains strategically important, but the old “persistent PR outlier” line is not currently citable after Born screening. The credible target is narrower: a magnetic mechanical-stability failure audit for Fe and related transition metals.
\nNext gate: compare spin-aware and spin-agnostic references under Born screening, phonon stability, elastic stability, and coupling-aware nulls.
\nKill if matched screened inputs remove the effect or if the signal is explained by one invalid tensor.
\nAu escape is a good hypothesis target, especially for catalyst-adjacent surfaces, but it remains open. The bulk pre-screening signal should become a surface/adsorbate canary with Ag/Pt/Pd controls.
\nNext gate: low-index slabs and simple adsorbates with MACE/CHGNet/ORB/SevenNet, reporting scalar errors plus error-subspace geometry.
\nKill if ORB/SevenNet do not reproduce the effect or if a group-level noble-metal explanation beats Au-specificity.
\nPhonons are strategically important because they expose second-derivative failures that scalar energy/force metrics can hide. Current repo evidence is mostly protocol and review, so this should be framed as the next canary protocol unless pilot measurements are added.
\nNext gate: a small sealed phonon subset across Si/Ge/diamond, MgO/Al2O3/SrTiO3, and graphite/MoS2/h-BN.
\nKill if reference functional mismatch dominates or displacement-size sensitivity changes the conclusion.
\nEvery top-priority target should ship as an evidence-carrying candidate, not as a press release:
\nThe next Lupine publication should not overclaim a new material. It should show the infrastructure that makes future material claims worth believing: a trust layer that catches negative transfer, refuses unsafe corrections, and publishes the evidence trail attached to each candidate.
\n"}