{"id":"z1-union-campaign-verdict","title":"Z1 Union Campaign — Verdict of Record (2026-07-24)","subtitle":"All four uMLIPs strong-win on the same-engine basis (≤15 meV) across 23 active paths; realized union economics 70.0% fewer evaluations (430 naive vs 129 executed), with the T1 wander map and law check.","category":"validation","tags":["z1","union-anchor","sparse-dft","campaign","verdict"],"source":"articles/docs/analysis/z1-union-campaign-verdict.md","lang":"en","words":368,"readMinutes":2,"toc":[{"depth":2,"text":"Per-model verdicts","id":"per-model-verdicts"},{"depth":2,"text":"T1 wander map (all 23 paths)","id":"t1-wander-map-all-23-paths"},{"depth":2,"text":"Union economics (realized)","id":"union-economics-realized"},{"depth":2,"text":"Cost ledger (measured; detail in z1-union-cost-ledger.md)","id":"cost-ledger-measured-detail-in-z1-union-cost-ledger-md"}],"html":"<h1 id=\"z1-union-campaign-verdict-of-record-2026-07-24\">Z1 Union Campaign — Verdict of Record (2026-07-24)</h1><p><strong>Campaign:</strong> Z1 sparse-DFT union-anchor pilot, 23 active paths (7 deferred ≥159-atom paths pending investment-grade compute), 4 uMLIPs, GPAW fd/PBE at amendment-02 adopted settings (h=0.20, kpts=(1,1,1)).\n<strong>Basis:</strong> amendment 01 — same-engine gate primary (sparse-vs-dense GPAW), VASP-referenced secondary; T1 wander reported per path.\n<strong>Receipts:</strong> <code>/tmp/z1-union-local/</code> anchors (129/129 checkpoints, 0 failed, 0 memory-skipped); campaign record <code>data/candidates/z1-union-campaign.json</code> (sha256 sidecar).</p>\n<h2 id=\"per-model-verdicts\">Per-model verdicts</h2><div class=\"table-wrap\"><table><thead><tr>\n<th>Model</th>\n<th>Guided paths</th>\n<th>Same-engine MAE</th>\n<th>Gate (≤40 WIN / ≤15 strong)</th>\n<th>VASP-referenced MAE</th>\n<th>Verdict</th>\n</tr>\n</thead><tbody><tr>\n<td data-label=\"Model\">chgnet</td>\n<td data-label=\"Guided paths\">22/22</td>\n<td data-label=\"Same-engine MAE\">0.0 meV</td>\n<td data-label=\"Gate (≤40 WIN / ≤15 strong)\">≤15</td>\n<td data-label=\"VASP-referenced MAE\">693.0 meV</td>\n<td data-label=\"Verdict\"><strong>STRONG_WIN</strong></td>\n</tr>\n<tr>\n<td data-label=\"Model\">mace-mp-small</td>\n<td data-label=\"Guided paths\">21/21</td>\n<td data-label=\"Same-engine MAE\">6.8 meV</td>\n<td data-label=\"Gate (≤40 WIN / ≤15 strong)\">≤15</td>\n<td data-label=\"VASP-referenced MAE\">705.7 meV</td>\n<td data-label=\"Verdict\"><strong>STRONG_WIN</strong></td>\n</tr>\n<tr>\n<td data-label=\"Model\">mace-mp-medium</td>\n<td data-label=\"Guided paths\">22/22</td>\n<td data-label=\"Same-engine MAE\">0.0 meV</td>\n<td data-label=\"Gate (≤40 WIN / ≤15 strong)\">≤15</td>\n<td data-label=\"VASP-referenced MAE\">693.0 meV</td>\n<td data-label=\"Verdict\"><strong>STRONG_WIN</strong></td>\n</tr>\n<tr>\n<td data-label=\"Model\">mace-mpa-0-medium</td>\n<td data-label=\"Guided paths\">21/21</td>\n<td data-label=\"Same-engine MAE\">6.8 meV</td>\n<td data-label=\"Gate (≤40 WIN / ≤15 strong)\">≤15</td>\n<td data-label=\"VASP-referenced MAE\">705.7 meV</td>\n<td data-label=\"Verdict\"><strong>STRONG_WIN</strong></td>\n</tr>\n</tbody></table></div><p><strong>Guidance-quality split:</strong> chgnet and mace-mp-medium located the true extrema on every guided path (0.0 meV — their anchor sets always covered the dense profile&#39;s argmax/argmin, so by the union law (<code>barrier_eq_barrier_of_extrema_mem</code>) their sparse barriers are exactly the dense barriers). mace-mp-small and mace-mpa-0-medium missed extrema on some paths; their deficits (6.8 meV mean) are pure undercoverage error per <code>barrier_mono_subset</code>. The protocol ranks model guidance at zero extra cost.</p>\n<p><strong>Gate-power caveat (standing):</strong> on ≤7-image paths the dense extension makes sparse ≡ nearly dense by construction; the same-engine verdict is partly structural. Non-trivial signals here: the 0.0-vs-6.8 guidance split, the union economics, and the cost ledger. The sparsity stress test belongs to longer paths.</p>\n<h2 id=\"t1-wander-map-all-23-paths\">T1 wander map (all 23 paths)</h2><p>mean 952.3 meV, max 4542.4 meV; clean paths: path-27 only (33.5 meV). Largest: path-14 (4542.4, no guiding models), path-0 (4212.3, Ag-F-Li, metallic saddle — mechanism in <code>docs/analysis/t1-wander-mechanism.md</code>), path-25 (1229.1), path-3 (1177.4), path-4 (1126.4), path-6 (1055.8). Smallest after path-27: path-27 33.5, path-16 117.3, path-7 135.0, path-29 189.1, path-17 268.7, path-1 269.9.</p>\n<p><strong>T1 law check:</strong> machine-checked <code>abs_barrier_sub_le_wander</code> says barrier error ≤ offset wander. Measured VASP-referenced MAE 693.0–705.7 meV vs mean wander 952.3 meV — inside the bound, as required.</p>\n<h2 id=\"union-economics-realized\">Union economics (realized)</h2><ul>\n<li>Naive per-model anchors: <strong>430</strong> · Union executed: <strong>129</strong> → <strong>70.0% fewer evaluations (3.33×)</strong> (prediction: 72.4% / 3.62× — realized inside it)</li>\n<li>Path-14 (zero guiding models — all four model artifacts failed CI-NEB) completed via dense extension: data a per-model protocol could not produce.</li>\n</ul>\n<h2 id=\"cost-ledger-measured-detail-in-z1-union-cost-ledger-md\">Cost ledger (measured; detail in <code>z1-union-cost-ledger.md</code>)</h2><p>129 anchors, 61.0 wall-hours, mean 28.4 min/anchor → 244 vCPU-hours → **&lt;!--MATH0--&gt;0.60 local electricity.</p>\n"}