Which Word in the Headline Has Been Measured?

On 4 October Vals AI published a post titled “Two Room-Temperature Antiferromagnetic Semiconductor Candidates.” Its author, Geby Jaff, says a team of Claude Opus 5.5 agents designed one new compound and, separately, found a second that chemists first made in 1999. By the next evening the Hacker News thread had passed 400 points (as of 6 October) under a livelier title that used the word “discover.”

I read the post, and then the public repository it points to, which Jaff calls a ledger. The repository is more careful than the headline. I took the headline one word at a time and asked what each has been checked against.

“Candidates”

This is the word the evidence fully supports. A candidate is a computed prediction, and the post says so in its first paragraph. The two predictions are YBaMnFeO5 and KV[Cr(CN)6], and both are claimed to be what the post calls Luttinger-compensated: zero net magnetic moment, yet electrons near the band edges sorted by spin, which is the property memory devices want and ordinary antiferromagnets lack.

The numbers come from density functional theory run by the agents. The slower, usually more accurate of the two methods used (HSE06) gives YBaMnFeO5 a band gap of 2.35 eV and KV[Cr(CN)6] about 2.1 eV, with spin-sorted windows of 1.0 to 2.6 eV, set against roughly 26 meV of thermal jiggle at room temperature. The ledger holds 876 included input files (868 pw.x calculations plus 8 post-processing runs), 61 numbered claims, and a checker that reports 58 passes, no failures, and 3 claims it cannot recompute. Code is MIT-licensed and data is CC BY 4.0. Most claims in this genre ship with less. The post frames its find against a 2025 study (Guo and colleagues) that predicted two other compounds of this type and found both lose magnetic order below room temperature, which is why a room-temperature one counted as the open goal.

“Room-temperature”

Here the two compounds part ways, and the figure below shows why.

Ordering temperatures, order-disorder temperature and synthesis window on one temperature axis Range chart on a temperature axis from 250 to 1650 kelvin with a reference line at 300 kelvin, room temperature. YBaMnFeO5 computed magnetic ordering temperature: 417 kelvin raw, 490 kelvin after calibration against a known magnet, and as low as 320 to 340 kelvin at a Hubbard U of 6 electronvolts. KV[Cr(CN)6] measured ordering temperature in a 1999 sample: 376 kelvin. YBaMnFeO5 manganese and iron order computed to dissolve at about 950 kelvin, 915 to 965 in the re-run. Oxide synthesis typically takes 900 to 1300 degrees Celsius, or 1173 to 1573 kelvin, entirely above the temperature at which the order is computed to dissolve. TEMPERATURES, KELVIN · VALS AI POST AND LEDGER, 4 OCT 2026 300 K, room 600 900 1200 1500 YBaMnFeO₅ orders KV[Cr(CN)₆] orders YBaMnFeO₅ Mn/Fe order Oxide synthesis computed, Monte Carlo measured, 1999 powder computed to dissolve typical oxide range 320 K, at U = 6 eV — ledger, 4 Oct 2026 417 K raw — ledger, 4 Oct 2026 about 490 K calibrated — Vals AI post, 4 Oct 2026 376 K measured — Holmes and Girolami, 1999, as reported by the post 915 K — ledger re-run, 4 Oct 2026 965 K — ledger re-run, 4 Oct 2026 1173 K, 900 °C — Vals AI post, 4 Oct 2026 1573 K, 1300 °C — Vals AI post, 4 Oct 2026 417 raw · 490 calibrated 376 about 950 1173 to 1573 Left dot of the first row: 320 K when the Hubbard U is raised to 6 eV, per the ledger.
The computed compound's magnetic order sits above room temperature on the post's own calibration, but the arrangement that produces it is computed to scramble about 200 K below the lowest temperature at which oxides of this kind are usually made, with a ledger uncertainty of +250/-150 K on that figure.

For KV[Cr(CN)6], “room-temperature” is a measurement, and an old one: the 1999 sample stayed magnetically ordered to 376 K, which is above 100 °C. That is the paper by Holmes and Girolami, whose title advertises exactly that. The measurement is of a powder with water in its pores, and it carried a small leftover moment, 0.125 Bohr magnetons per formula unit where a perfect crystal would have none. The two computational methods disagree about what the water does: one says the spin sorting survives, the other says the hole window shrinks by more than half.

For YBaMnFeO5, “room-temperature” is a Monte Carlo estimate: 417 K raw, about 490 K after the post calibrates against a known magnet, and 320 to 340 K if the Hubbard U (a correction knob in the calculation) is set to 6 eV instead. The post then concedes the larger problem. The design needs manganese and iron to sit in a perfect checkerboard, the agents’ simulation has that order dissolving near 950 K, and oxides of this type are made at 900 to 1300 °C. The ledger’s re-run puts the dissolving temperature at 915 to 965 K. It gives the headline figure as 950 K with an uncertainty of +250/-150 K, calls its cluster expansion provisional, and notes that its highest variant (1210 K) reaches into the synthesis window. The post’s own reading is that standard synthesis would likely give a scrambled crystal that loses the spin sorting. So the headline’s room-temperature compound may be hard to make in the form the calculation describes.

“Semiconductor” and the spin sorting

The gaps and windows are computed for ideal, zero-kelvin crystals. The post says that neither the band gap nor the spin sorting of KV[Cr(CN)6] has been measured, and I found no measurement of either in anything the ledger or the post cites. The ledger’s own caveat list runs to seventeen items. Among them: compensation is generally exact only at zero temperature, so a net moment can appear when it warms; the HSE06 windows come from a coarse k-point grid; linear-response U values were never finished; and heavy carriers and self-trapping are expected, so transport should not be silicon-like. A different pseudopotential family gives a PBE+U gap of 1.795 eV against 1.953 eV, which the ledger labels a robustness check, not a reproduction.

What the re-runs bought

How the ledger's 61 claims were checked Bar chart of the ledger's 61 numbered claims by how each was checked. 52 were recomputed from raw outputs, 3 by re-running scripts, 3 read from analysis files, and 3 are literature or experimental values. The checker reports 58 passes, no failures, and 3 claims it cannot compute. LEDGER CLAIMS BY CHECK METHOD · 61 CLAIMS · LEDGER, 4 OCT 2026 Recomputed from raw outputs Re-ran a script Read from an analysis file Literature or experiment 52 claims — ledger, 4 Oct 2026 3 claims — ledger, 4 Oct 2026 3 claims — ledger, 4 Oct 2026 3 claims — ledger, 4 Oct 2026 52 3 3 3
The ledger's checker reports 58 passes, no failures and 3 claims it cannot compute. Every computed check was run by the author's own pipeline; the 3 literature values were not recomputed.

The ledger says four sets of re-runs reproduced 16 checked values, though its RESULTS.md tabulates 14 rows across three reproduction sets plus the pseudopotential robustness check. I believe the matches. But “independent” here means a fresh container and checked pseudopotential files, run through the author’s own pipeline and not by an outside group. That verifies that the arithmetic repeats. It does not test whether PBE+U or HSE06 is the right physics for these two compounds. On that, the ledger says the agents’ preference for HSE06 is “a judgement, not a measurement,” and that PBE+U gets the ground state of a twin compound, YBaMn2O5, wrong at U of 2 eV or higher.

The ledger also lists five corrections to its own earlier numbers. The YBaMnFeO5 distance above the stability hull went from +2.6 to +13.7 meV per atom once two competing compounds were added, and an unconverged HSE06 run for the hydrated crystal was finished, moving the hole window from 2.43 to 2.31 eV. A correction list with numbers on it is the thing I would want more groups to publish. It also shows how much of the reported precision depended on individual runs finishing, though the ledger says none of the five corrections changes its main conclusions.

The word “discover”

The post itself uses “designed” for the new compound and “identified” for the old one. The second case is the more interesting. The 1999 chemists made KV[Cr(CN)6] so that its two metals’ moments cancel. A 2008 hybrid-functional study of Prussian blue analogues at pressure, by Middlemiss, Lawton and Wilson, plotted its electron states spin by spin, and the ledger concedes that same-spin band edges are visible in that plot, unremarked. The agents’ contribution was to compute the spin-sorted windows and name the property, not to spot the 2008 plot: the ledger records that the agents’ own literature search missed that paper.

Noticing is work. The old Prussian-blue-family compound had been in the literature for 27 years. The idea of a magnet with zero net moment and spin-polarized carriers was published as theory in 1995, when van Leuken and de Groot proposed half-metallic antiferromagnets in Physical Review Letters. The repository’s README gives the grades from its agents’ adversarial review: “design study, not a realizable discovery” for YBaMnFeO5 and solid “identification plus numbers,” not a breakthrough, for KV[Cr(CN)6]. That is a fair description, and well below the Hacker News title.

I wrote earlier about what it takes to confirm a robot-made material: proposing candidates was never the bottleneck, identifying what came out of the furnace was. This case sits one step earlier in the same chain. The next step is the one the post names itself, which is to make KV[Cr(CN)6] again and measure it, with the saturation moment, spin-resolved photoemission or X-ray magnetic circular dichroism at the vanadium and chromium edges as the tests the ledger proposes. I found no synthesis or measurement claim as of this morning, two days after posting, which is unsurprising two days in; the interesting part has not happened yet.

References

  1. Jaff, G. (2026). Two Room-Temperature Antiferromagnetic Semiconductor Candidates. Vals AI, 4 October 2026. The calculated values, the Tc figures and the synthesis-temperature range are the company’s own.
  2. Jaff, G. (2026). compensated-magnet-ledger. GitHub repository, commits of 4 October 2026 (README, LEDGER.md, reproduce/RESULTS.md). Counts of runs, claims, re-runs and corrections are taken from it; its caveat list has 17 numbered items.
  3. Holmes, S. M., Girolami, G. S. (1999). Sol-Gel Synthesis of KV[Cr(CN)6]·2H2O: A Crystalline Molecule-Based Magnet with a Magnetic Ordering Temperature above 100 °C. Journal of the American Chemical Society 121(23):5593–5594, June 1999. The 376 K figure is as reported in the Vals AI post and ledger; I could not open the paper itself.
  4. Middlemiss, Lawton, Wilson (2008). A solid-state hybrid density functional theory study of Prussian blue analogues and related chlorides at pressure. Journal of Physics: Condensed Matter 20, 335231. Cited through the ledger; not read directly.
  5. van Leuken, H., de Groot, R. A. (1995). Half-Metallic Antiferromagnets. Physical Review Letters 74:1171–1173. Background.
  6. Guo, Hou, Gao, Yang, Ji, Lu (2025). Luttinger compensated bipolarized magnetic semiconductor. arXiv:2502.18136, 25 February 2025; Physical Review B 112, November 2025. The study the post cites, which it says predicted two Luttinger-compensated semiconductors that lose magnetic order below room temperature. Background.
  7. Hacker News (2026). Opus 5.5 agents discover two room-temperature magnetic semiconductor candidates. Thread, 5 October 2026.