The most revealing number in a new AI materials report is zero. Neither of its two magnetic-semiconductor candidates has had its predicted spin polarization, band gap or conductivity measured. The report, which reached 316 points on Hacker News, describes a Claude Opus 5.5 agent team designing one compound and recognizing a useful property in another. Its own evidence draws a harder boundary: the agents have produced targets for a laboratory, not confirmed materials discoveries.
That distinction matters because the headline result is easy to flatten into a story about AI doing science by itself. Vals AI says several agents ran density functional theory calculations on cloud computers between October 1 and 4, while a human set the goals, directed the search and decided what to publish. One candidate exists only as a computer model. The other was synthesized 27 years ago, and part of its electronic behavior appeared in a 2008 calculation that the agents initially missed. The project's public computational ledger records both the promising numbers and the corrections that narrowed the claim.
What the agents were looking for
The search focused on Luttinger-compensated magnetic semiconductors. In an ordinary ferromagnet, atomic magnetic moments point together, creating a net field and separating electrons by spin. In an antiferromagnet, neighboring moments point in opposite directions and cancel. A Luttinger-compensated material aims for a useful combination: opposing moments add up to zero in an ideal crystal while inequivalent atomic sites still separate electronic states by spin. The Vals explanation frames that combination as attractive for spin-based memory because low stray fields could allow close packing while the spin-separated states remain readable.
The size of the spin window is central to the claim. It is the energy range at a band edge containing states of only one spin. Room-temperature thermal motion is about 26 millielectronvolts, according to the project. Its HSE06 calculations give much larger windows for both candidates: 1.0 and 1.4 electronvolts for holes and electrons in YBaMnFeO5, and 2.6 and 1.6 electronvolts for KV[Cr(CN)6]. The same calculations estimate band gaps of 2.35 eV and 2.09 eV. Those figures come from idealized crystal models, as the ledger definitions and run settings make clear.
There was already a stated target in the literature. A 2025 paper on Luttinger-compensated semiconductors predicted Mn(CN)2 and Co(CN)2, but calculated magnetic ordering temperatures below room temperature. Its authors called for candidates that combine a semiconductor-scale gap with magnetic order above room temperature. The agent search appears to have aimed directly at that missing combination.
The fresh design failed its practical test
YBaMnFeO5 is the new compound in the report. In the desired structure, manganese and iron occupy an exact checkerboard arrangement. The simulations predict zero net spin moment, the 2.35 eV band gap and magnetic order around 420 kelvin before calibration. A comparison with known magnets pushed the calibrated estimate to roughly 490 K. On those calculations alone, the material looks like the cleaner of the two candidates in the project's results table.
Its problem appears when the calculation moves closer to synthesis. The agent team modeled how the manganese and iron atoms would arrange at different temperatures and found the required checkerboard breaking into a random mixture at roughly 950 K. The report says ordinary synthesis of this oxide would take about 900 to 1,300 degrees Celsius. At lower temperatures, the atoms barely move. Standard preparation could therefore lock in a scrambled arrangement, and one simulated manganese-iron swap was enough to close the spin-selective gap.
The repository accordingly grades YBaMnFeO5 as a design study rather than a realizable discovery. That negative result is scientifically useful. It stops an appealing set of electronic calculations from becoming a vague claim that someone has found a manufacturable material. A chemist could still explore an unusual low-temperature or nonequilibrium route, but the published work offers no demonstrated synthesis path. The repository's experiment notes put the burden where it belongs: on making the ordered crystal before discussing a device.
The old powder is the better lead
KV[Cr(CN)6] has the opposite profile. Stephen Holmes and Gregory Girolami made a hydrated form in 1999 using a sol-gel process. Their Journal of the American Chemical Society paper reported magnetic ordering at 376 K, or 103 degrees Celsius. Repeated heating reduced that figure to 365 K, which the researchers linked to dehydration. The sample also had a small residual magnetic moment rather than perfect cancellation, plausibly because its composition and vanadium oxidation state were not ideal.
The Claude agents' contribution is a new reading of that material. Their calculations predict that an ideal, dry crystal has zero net spin moment, a 2.09 eV band gap and same-spin states at both band edges. The Vals report says a 2008 hybrid-functional study had already plotted the same-spin band edges without identifying the compound as a Luttinger-compensated semiconductor. So the agents did not uncover a forgotten physical sample or originate every relevant calculation. They connected earlier experimental and computational evidence, quantified the spin windows and tested defects that prior work had not treated as the main question.
Water is the unresolved complication. The 1999 material was a hydrated powder, while the headline electronic predictions describe a perfect dry crystal. When the team included water, its two computational methods disagreed. HSE06 preserved wide spin windows after a corrected run converged, while PBE+U cut the hole window by more than half. The ledger also records a water-filled vacancy producing a nonzero magnetic moment. No experiment has yet measured whether the real powder has the predicted spin-sorted band edges.
The audit trail is part of the result
The agent workflow is easier to assess than many AI research claims because it left behind raw calculations, scripts and failed checks. The repository indexes 876 Quantum ESPRESSO inputs, including 868 pw.x calculations. Its checker recomputed 58 numerical claims from the supplied files with zero reported failures on October 4, while three literature or experimental values could not be recomputed. Four sets of calculations were also rerun in a fresh cloud container, where 16 checked values matched the originals, according to the reproduction record.
That record does not make the scientific claims self-validating. It does make mistakes visible. During review, the team found that the calculated stability distance for YBaMnFeO5 had used an incomplete set of competing phases. The corrected value was 13.7 meV per atom rather than 2.6. One HSE06 water calculation had been marked converged even though its exact-exchange loop stopped early. A literature search missed the 2008 work. The corrections log keeps each issue beside the final claims instead of erasing the route that produced them.
Several limits remain outside the ledger's arithmetic. The calculations model ideal crystals at zero temperature. They omit reliable treatment of spin-orbit coupling and orbital moments, which may leave a residual magnetic signal. Density functional methods can place energy levels off by tenths of an electronvolt. Matching a stored output also verifies that a calculation can be reproduced, rather than proving that its model describes a real sample. Those qualifications are listed by the authors themselves.
For AI-assisted science, that may be the more transferable part of the work. Parallel agents searched candidate space, ran standard physics software and assigned adversarial reviewer agents to attack the claims. The process still missed prior art and initially accepted an unfinished calculation. Pre-registered pass or fail rules, retained raw outputs and human publication judgment made those errors recoverable. The repository documents roughly 750 jobs in this search lane and says other lanes that failed their thresholds were left out, which also means readers cannot measure the full search's hit rate from the published subset alone. The production notes state that limitation plainly.
The next result belongs to a laboratory
The clearest next step is to remake KV[Cr(CN)6], measure its composition and check whether its opposing magnetic sublattices nearly cancel. After that come element-specific magnetic measurements and spin-resolved photoemission, which could directly test the predicted spin polarization near the valence-band edge. The project lists those experiments in order and gives the photoemission prediction: about 100 percent one spin across roughly 2 eV in the dry material. Until such measurements exist, the candidate remains a calculation attached to an old powder.
Watch whether an independent lab can reproduce the 1999 compound and measure the electronic behavior the agents predict. The project's proposed tests make that result legible: a confirmed spin window would turn the old material into the stronger scientific result, while a failed test would show where the models broke. Either outcome would move this story beyond a 316-point discussion and give the agent workflow the one check its own files cannot supply: contact with matter.