\n\nVersion: 1.0 (2026-07-01)\nObjective: Determine whether foundation-MLIP error geometry spans property\nfamilies — i.e., whether the low-dimensional structure previously measured\ninside elastic constants is a property of the model (a real ribbon in\nprediction space) or an artifact of the property family's internal physical\ncoupling (Cauchy relation, stability constraints).\nRegistered before confirmatory analysis of the first cross-property sweep\n(
\ndata/y_matrix_runs/, launched 2026-07-01). The sweep's machinery smoke\ntests (Ni/MACE-MP-small) were observed during development and are quarantined\nas exploratory — see §Exploratory quarantine.
| Date | \nChange | \nAuthor | \n
|---|---|---|
| 2026-07-01 | \nv1.0 — initial registration | \nsquad (Claude + Alex Welcing) | \n
All prior evidence in the ledger (~1,320 records) sits inside one property\nfamily: 0 K cubic elastic constants. Elastic constants are physically coupled\nto each other, so low-dimensional error structure measured within them cannot\ndistinguish "the model's error manifold is a hyper-ribbon" from "elasticity is\ninternally coupled." The Y-matrix adds property families with different\nphysical couplings — point defects, surfaces, planar faults, equations of\nstate, formation enthalpies — which breaks the degeneracy. Whatever the\noutcome, the correction program ("the turbo") ships; this experiment determines\nits gearing (one shared correction operator vs per-family operators).
\nPer material, as structurally applicable:
\n| Family | \nProperties | \nUnit | \n
|---|---|---|
| Lattice/cohesion | \na₀ | \nÅ | \n
| EOS | \nB₀, B₀′ | \nGPa, — | \n
| Point defect | \nE_vac | \neV | \n
| Surfaces | \nγ₁₀₀, γ₁₁₀, γ₁₁₁ (fcc); γ₁₀₀, γ₁₁₀ (bcc) | \nJ/m² | \n
| Planar fault | \nγ_SFE (fcc only) | \nmJ/m² | \n
| Compound stability | \nΔH_f (intermetallics) | \neV/atom | \n
Computed by python/lupine_distill/statics/ (226-test suite; bit-reproducible\n≤1e-11), CLI python/scripts/run_y_matrix_statics.py, GPU lane pinned in\npython/requirements-gpu-lane.lock.
Explicit fail-fast model registry — no silent substitution (registered as a\ndesign invariant after the torchsim model_id defect, fixed 2026-07-01).\nExpansion models (M3GNet/TensorNet via matgl, Orb, SevenNet, UMA) join under\nthe same protocol when their lanes are pinned.
\nReference values live exclusively in data/y_matrix_targets/*.json\n(schema lupine.y_matrix_targets.v1), compiled under a no-fabrication\nrule: every value carries a verified citation; unverifiable values are\nrecorded as unresolved gaps, and properties whose reference is unresolved\nare excluded from confirmatory hypothesis tests (they may be reported\ndescriptively, labeled as such). Where DFT and experimental references both\nexist they are separate entries; the confirmatory target per property is the\nDFT-PBE value when available, else experiment, chosen now to match the\ntraining-data provenance of the model grid (all MP/MPtrj-derived). Elastic\nreference values additionally require resolution of the known\nFamily A/B discrepancy (audit finding, task #11) before elastic properties\njoin the cross-property vectors.
For each model, assemble per-material error vectors across all\nreference-resolved properties (normalization: signed relative error\n(pred − ref)/|ref| per property; properties with |ref| below 10× its\ncompilation uncertainty use absolute error scaled by the family's median |ref|\n— guard against near-zero-reference degeneracy, e.g. SFE). The participation\nratio of the material-by-property error covariance is lower than the 95th\npercentile of the coupling-aware null (§Nulls).
\nKill: PR within or above the null band → the elastic-era "ribbon" is not a\ncross-property object; report as refutation of cross-property universality.
\nThe leading principal error mode (from H1's covariance) has pairwise cosine\nsimilarity > 0.7 across the three models, exceeding the 95th percentile of\nthe same statistic under the null.
\nKill: similarity indistinguishable from null → error geometry is\nmodel-specific; correction operators do not transfer across architectures.
\nProperties underrepresented in MPtrj-style training data (E_vac, γ_SFE,\nsurfaces) show larger median |relative error| than bulk-adjacent properties\n(a₀, B₀, ΔH_f) for every model, by at least a factor of 2.
\nKill: any model where defect-family and bulk-family errors are comparable\n(< 1.5×) → the weak-spot story is model-dependent, not distributional.
\nThe elastic-trained scalar-bulk operator (v0.2, the Round-2 win), applied to\nY-matrix cells, reduces cross-property error norm. Staged: the operator\nimplementation is not currently in this repo (recovery/reimplementation\npending); H4 binds when it lands, before its transfer analysis runs.
\nKill: operator improves elastic but degrades ≥ half of non-elastic\nproperties → correction is family-local; the turbo ships with per-family\ngearing.
\nContext: H1/H2 were killed and H3 passed (see\ny-matrix-confirmatory-results-2026-07-01.md), so the registered expectation\nis now that transfer FAILS; observing transfer would be the surprising outcome.
s = median(B0_ref / B0_pred) fitted leave-one-material-out over the\nprimary matched set (15 metals; the held-out material never contributes to\nits own correction).s × prediction for\nenergy/stiffness-like properties (B0, E_vac, gamma_100/110/111, gamma_SFE,\ndH_f); a0 is excluded from correction (a length, not an energy scale — the\nsoftening mechanism predicts no first-order a0 effect).Published claim (Deng et al., npj Comput. Mater. 2024/25, arXiv:2405.07105;\ncorroborated arXiv:2410.12771): softening is data-distribution-driven — models\ntrained with off-equilibrium data (OMat24 lineage) largely eliminate it.\nRegistered prediction: MACE-MPA-0 (medium-mpa-0), run through the identical\n63-cell matrix and H3 analysis, shows a defect/bulk median-error ratio less\nthan HALF of MACE-MP-medium's (i.e., < 7.66 vs 15.32 on the primary matrix),\nand its per-model B₀ softening scalar s sits closer to 1 than any MPtrj model's.
\nKill: MPA-0 ratio ≥ half of MACE-MP-medium's → distributional retraining\ndoes not collapse the defect/bulk split on elemental metals; the contested\nsymmetry-related residual story (arXiv:2507.15190) gains weight.\n(Citation correction 2026-07-02, post-verdict: arXiv:2507.15190 is withdrawn\nand its abstract never made the attributed claim; the kill's interpretation\nclause should read simply "a residual beyond distributional bias exists,"\nsupported by our own R2 decomposition only.)\nSame seed/nulls/exclusions as Round 1; Fe deficiency annotation carries over\npending its per-model shape diagnostic.
\nNaive permutation nulls lie when properties are internally coupled. The\nregistered null: within-family structure preserved, cross-family alignment\nbroken — for each property family independently, permute the material labels\nof that family's error sub-vectors (1,000 draws; matched-n; per-material\ncompleteness required — materials missing any confirmatory property are\nexcluded from the vector analysis rather than imputed). This preserves each\nfamily's internal covariance (including elastic self-coupling) and destroys\nonly the cross-family correlation that H1/H2 claim.
\nObserved during machinery development, before this registration, on\n(Ni, MACE-MP-small): E_vac ≈ 1.02 eV vs ~1.45 eV PBE; γ_SFE ≈ −8 mJ/m² vs\n~145 mJ/m² PBE; a₀/B₀/ΔH_f near reference. These observations motivated H3 and\nare excluded from H3's confirmatory test for that (material, model) cell;\nH3 is evaluated on the untested cells.
\nEvery cell emits lupine.mlip.calc_evidence.v1 (canonical-inputs sha256,\nmodel id as-run, per-property values) to data/y_matrix_runs/; reference\nannotation and Lean emission (emit_lean_module, within/exceeds encoded\neither way) follow reference binding. Ledger landing is gated on the ingest\nrange fix (task #12) and does not gate the analysis.
| Date | \nDeviation | \nReason | \n
|---|---|---|
| 2026-07-01 | \nFe excluded from confirmatory vectors pending magnetic-state verification. | \nSweep Fe/bcc B₀ came back 71–89 GPa across all three models (vs ~170 GPa experimental) — consistent with the calculators relaxing to the nonmagnetic branch because statics.structures sets no initial magnetic moments. This is a state-preparation issue, not model error; attributing it to the models would be a false negative-transfer claim. Fe rejoins after initial-magmom support is added and cells re-run. Logged before reference binding. | \n
| 2026-07-01 | \nH1 criterion wording defect, logged at analysis time: the registered sentence "lower than the 95th percentile of the coupling-aware null" is vacuous as written; the analysis implements the intended strict one-sided test PR < null p05 and reports both readings. See y-matrix-confirmatory-results-2026-07-01.md §Deviations. | \nWording error in v1.0; the strict reading is adverse to the hypothesis (H1 passes 2/6 under it vs ~6/6 under the vacuous literal reading), so the clarification cannot inflate the result. | \n
| 2026-07-01 | \nAMENDMENT (same day, still before reference binding): Fe re-admitted to confirmatory vectors with a documented-model-deficiency annotation; the previous row's reason is empirically falsified. | \nCPU investigation (scratchpad fe_investigation/memo.md): initial magnetic moments change the energy of neither calculator (ΔE = 0.0 eV exactly; MACE is spin-blind, CHGNet magmoms are outputs — it predicts FM Fe at 2.44 μB and is soft anyway). B₀ ≈ 70–90 GPa is fit-window-independent (±3% → ±15%: <1 GPa change for MACE-small) with a curvature anomaly (B₀′ = −12.9/−7.4 in the stored fits) matching the published PES-softening pathology of MPtrj-trained uMLIPs on magnetic Fe (Deng et al. arXiv:2405.07105; Springer J. Phase Equilib. Diffus. 2025; arXiv:2605.28395). The "magmom support + re-run" remedy is struck (provable no-op). Fe is therefore genuine negative-transfer evidence, exactly what the Y-matrix measures. Protocol additions: BM3 fits auto-flagged when B₀′ ∉ [1, 8]; magnetic elements (Fe, Cr, Mn, Co, Ni) get a one-shot wide-scan shape diagnostic before binding. Cr is NOT pre-excluded (its sweep-2 cell shows the opposite symptom: stiff, B₀′ normal). | \n