Status: self-corrected / open gates\nPublication date: 2026-06-20\nScope: foundation-MLIP elastic tensors, the classical-to-MLIP transfer conjecture, Fe outlier language, and Au escape language.
\nThe earlier MLIP-transfer story was too compressed: we had pre-screening counts that made the classical-to-MLIP hyper-ribbon look like a clean “14 of 15 elements stay on the ribbon” result, with Fe as the stable exception and Au as an escape event. A later Born-stability screen found that 7 of 45 foundation-model elastic tensors were mechanically invalid, including CHGNet-Fe plus MACE-V and Orb-v3 Al/Nb/Pb/Pt.
\nThat does not erase the program. It narrows what can be cited.
\nWhat survives now is the more defensible claim: after screening, cross-model error axes remain strongly one-dimensional in the available Born-stable inputs, with rank-1 share about 0.56–0.94 at n = 8–11 models per element. What does not survive as citable evidence is the old per-element “14/15 on-ribbon” count, the old “Fe PR > 2 under every LAM addition” line, or any figure that depends on unscreened foundation-model elastic tensors.
\nThe library is meant to be a claim ledger, not a highlight reel. A good discovery system must publish the moment when an attractive result becomes less simple. The re-audit changes the MLIP-transfer shelf from “confirmed count” to “open recomputation with a surviving directional signal.” That is a stronger public position than keeping the old headline and burying the caveat.
\n| Claim | \nPrevious public shorthand | \nCurrent citable status | \n
|---|---|---|
| Classical hyper-ribbon universality | \nSupported | \nSupported for the classical OpenKIM/NIST-style elastic corpus. | \n
| Classical → MLIP transfer count | \n14/15 elements stay on-ribbon under MACE/CHGNet/Orb-v3 | \nOpen. Pre-screening count is frozen until recomputed on Born-stable inputs. | \n
| Fe persistent outlier | \nFe remains PR > 2 across the foundation-MLIP trio | \nOpen / narrowed. CHGNet-Fe failed Born stability; Fe remains a magnetic failure-mode suspect, not a citable PR exception. | \n
| Au escape | \nAu escapes under MACE and CHGNet; Ag escape refuted | \nOpen. Pre-screening signal is useful for hypothesis generation; mechanism and screened robustness remain unproven. | \n
| Cross-model directional structure | \nLow-dimensional error axes appear across MLIPs | \nSupported as a directional result in screened inputs, not yet a settled per-element count. | \n
The current ledger note records that Born screening excluded 7 of 45 foundation-model elastic tensors, including CHGNet-Fe, MACE-V, and Orb-v3 Al/Nb/Pb/Pt. On the surviving inputs, the error axis remains one-dimensional for all 15 elements in the re-audit window, with rank-1 share 0.56–0.94. Fe’s cross-model alignment is about +0.80, which means it is not currently an alignment outlier.
\nThe resulting scientific interpretation is narrower and cleaner: invalid elastic tensors were themselves a meaningful failure mode, but they cannot be used as evidence for participation-ratio counts.
\nOlder figures and reports that show Au PR jumps, Fe PR > 2, or 14/15 MLIP counts are historical/pre-screening artifacts unless they explicitly say otherwise. They remain useful as method demonstrations: how to compute participation ratio, how an escape detector is supposed to behave, and how a claim moves through the ledger. They should not be quoted as current evidence for screened MLIP behavior.
\nThe next credible publication gate is a Born-stable recomputation:
\nThe classical-to-MLIP transfer count should remain open, or be downgraded, if any of these hold:
\nThe honest claim is no longer “foundation MLIPs preserve the ribbon for 14 of 15 elements.” The honest claim is:
\n\n\nBorn screening invalidated the old MLIP transfer count. The low-dimensional directional structure still appears to survive in screened inputs, but the per-element lifecycle status is open until recomputed.
\n
That is the version the library should carry forward.
\n"}