Status: Open — under Born-screened re-audit
\nThe pre-screening “14/15 elements stay on the ribbon” count is no longer citable\nas a settled MLIP-transfer result. Born stability screening (see\nreplication/error-geometry/, prereg commits dffbe595/ebf39e33) excludes 7\nof 45 foundation-model elastic tensors — including CHGNet-Fe (C11 < |C12|),\nMACE-V, and Orb-v3 Al/Nb/Pb/Pt — so any per-element participation-ratio or\non-ribbon count computed from the unscreened trio inputs must be recomputed.
What survives is narrower but still important: at n = 8–11 Born-stable models per\nelement, the error axis remains strongly one-dimensional for all 15 elements\n(rank-1 share 0.56–0.94). The current claim is therefore directional\nlow-dimensional structure in screened inputs, not a restored 14/15 per-element\ncount.
\nThe hyper-ribbon is not a quirk of classical force fields. The live question is\nwhich parts of that low-dimensional error geometry survive when modern foundation\nmachine-learning interatomic potentials (MLIPs) are added to the corpus after\nmechanical-stability screening.
\nIngested one foundation MLIP at a time on the 15 IMMI elements:
\nSeparately, the corpus was de-myopized: it had been ~99.5 % elastic-constant\nrecords. Recovering real lattice constants (a0) from MLIP provenance (45 records)\nand forcing a joint Cij+a0 manifold, the ribbon survives (participation\nratio 1.05–2.05). So it is not an artifact of one property family.
\nThis remains the genuinely important target: the prior — stated explicitly before\nthe test — was that the ribbon/dichotomy would not transfer from classical to\nMLIP. The screened directional signal suggests the geometry may still transfer,\nbut the old element count cannot be promoted until it is recomputed on valid\ntensors.
\nFirst recompute the classical-to-MLIP transfer count on Born-stable inputs. Then\nadd Ecoh and B0 from the Phase-D compute pipeline so the manifold\nspans four property families, and re-test ribbon stability across the wider basis.
\n"}