{"id":"hyp-cross-mlip-orthogonal-errors","title":"Cross-MLIP Orthogonal Error Modes","subtitle":"MACE and CHGNet fail in orthogonal directions — the precondition for ensembling.","category":"conjectures","tags":["mlip","ensemble","error-geometry"],"source":"articles/docs/conjectures/cross-mlip-orthogonal-errors.md","lang":"en","words":178,"readMinutes":1,"toc":[{"depth":2,"text":"Claim","id":"claim"},{"depth":2,"text":"Evidence","id":"evidence"},{"depth":2,"text":"Why it matters","id":"why-it-matters"},{"depth":2,"text":"Next","id":"next"}],"html":"<h1 id=\"cross-mlip-orthogonal-error-modes\">Cross-MLIP Orthogonal Error Modes</h1><p><strong>Status: Supported</strong></p>\n<h2 id=\"claim\">Claim</h2><p>Different foundation MLIPs do not just have <em>different-sized</em> errors — they have errors\npointing in <strong>orthogonal directions</strong> for certain elements. The mistakes are\nstructured and model-specific, not shared noise.</p>\n<h2 id=\"evidence\">Evidence</h2><ul>\n<li>MACE-MP-0 and CHGNet have <strong>orthogonal error directions on Ag, Nb, Pd</strong> — where one\nmodel is wrong, the other is wrong differently, not more.</li>\n<li>This is what lets CHGNet &quot;pull back&quot; Ag while MACE lets it escape (see\n<a href=\"#/read/hyp-au-mlip-escape\">Au/Ag escape</a>) — the models disagree in direction, not magnitude.</li>\n<li>The cross-MLIP alignment hypothesis (<code>hyp_mlip_alignment_test</code>) — that the\nelement-intrinsic error dichotomy would <em>extend</em> to foundation MLIPs as a shared\naxis — was <strong>refuted</strong> (ρ = 0.19, p = 0.51). Orthogonality is the surviving picture.</li>\n</ul>\n<h2 id=\"why-it-matters\">Why it matters</h2><p>Orthogonal error modes are exactly the precondition for <em>ensemble</em> error correction:\nif two models fail independently, combining them is informative. A shared error axis\nwould have made ensembling useless. This is the constructive flip-side of the\nalignment refutation.</p>\n<h2 id=\"next\">Next</h2><p>Quantify the orthogonality as an ensemble-gain estimate: how much error reduction is\navailable from a MACE+CHGNet+Orb ensemble on the orthogonal elements?</p>\n"}