{"id":"mlip-ni-paired-accuracy-live","title":"Ni Paired Accuracy Live Report","subtitle":"Live baseline-versus-Distill Accuracy evidence for the Ni fcc EAM-home-turf campaign.","category":"validation","tags":["mlip","distill","ni","cloud-run","live-lab"],"source":"articles/docs/mlip-ni-paired-accuracy-live-report.md","lang":"en","words":524,"readMinutes":2,"toc":[{"depth":2,"text":"Why This Run Exists","id":"why-this-run-exists"},{"depth":2,"text":"Live Summary","id":"live-summary"},{"depth":2,"text":"Current Interpretation","id":"current-interpretation"},{"depth":2,"text":"Flagship Promotion Gate","id":"flagship-promotion-gate"},{"depth":2,"text":"Pair Table","id":"pair-table"},{"depth":2,"text":"Evidence Contract","id":"evidence-contract"},{"depth":2,"text":"Release Read","id":"release-read"}],"html":"

Ni Paired Accuracy Live Report

Status: rejected candidate: negative transfer detected\nGenerated: 2026-05-27T12:58:20Z\nCampaign: ni-fcc-eam-home-turf-paired-accuracy-v1\nFixture hash: sha256:0a51d5b58f68e169478b290954da4729e33d9d37c2b65a3084e60ad5c1edd16b

\n

Why This Run Exists

The previous 25-cell baseline showed that our cloud MLIP runner can score five\nmodels across five physics rows. This run asks a stricter question: can Lupine\nDistill Accuracy improve the same MLIP on the same sealed Ni fcc EAM-home-turf\nfixture while sharing raw-prediction evidence with the paired baseline?

\n

That pairing matters. For each (row, MLIP), the baseline cell writes the raw\nprediction checkpoint and the Distill cell reads that same checkpoint. A claimed\naccuracy lift therefore has to come from the Distill policy layer, not a changed\nmodel invocation, changed fixture, or untracked rerun.

\n

Live Summary

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
MeasureValue
Total cells50
Completed cells50
Failed cells0
Missing cells0
Total paired comparisons25
Measured paired comparisons25
Improved pairs0
Regressed pairs10

Current Interpretation

At this snapshot, 25 paired comparisons have both baseline and\nDistill artifacts. 0 improve, 10 regress, and the\nremaining rows are awaiting artifacts or explicit failures. No missing cell is\ntreated as a win.

\n

The returned evidence is currently a negative-transfer finding: 25 pairs are measured, 10 regress, 15 are unchanged, and none improve. That is still useful because the system has caught a ribbon that should refuse or adapt before it is promoted for this material lane.

\n

Flagship Promotion Gate

Result: blocked. This campaign is evidence for ribbon rejection, not launch promotion.

\n\n

Required next action: reject this ribbon for flagship claims; fit a material-family-aware canary and require zero regressions before rerun

\n

Pair Table

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
RowMLIPBaseline errorDistill errorLiftVerdict
Elastic constantschgnet27.71127.7110.0%unchanged
Elastic constantsm3gnet30118.0430118.040.0%unchanged
Elastic constantsmace-mp-022.65322.6530.0%unchanged
Elastic constantsorb-v38.00018.00010.0%unchanged
Elastic constantssevennet26.74026.7400.0%unchanged
Energy-volumechgnet1.29431.3005-0.5%distill regressed
Energy-volumem3gnet1.28771.4476-12.4%distill regressed
Energy-volumemace-mp-01.28031.4168-10.7%distill regressed
Energy-volumeorb-v31.04381.2133-16.2%distill regressed
Energy-volumesevennet1.31511.3696-4.1%distill regressed
Forceschgnet0.10540.1054-0.0%unchanged
Forcesm3gnet0.04870.0487-0.0%unchanged
Forcesmace-mp-00.08250.08250.0%unchanged
Forcesorb-v30.06840.06840.0%unchanged
Forcessevennet0.03910.03910.0%unchanged
Relaxation stabilitychgnet1.29301.2943-0.1%distill regressed
Relaxation stabilitym3gnet1.28651.3968-8.6%distill regressed
Relaxation stabilitymace-mp-01.27981.4025-9.6%distill regressed
Relaxation stabilityorb-v31.04271.2328-18.2%distill regressed
Relaxation stabilitysevennet1.31381.4340-9.1%distill regressed
Stresschgnet1.44791.44790.0%unchanged
Stressm3gnet285.89285.890.0%unchanged
Stressmace-mp-00.86110.86110.0%unchanged
Stressorb-v31.13241.13240.0%unchanged
Stresssevennet1.48461.48460.0%unchanged

Evidence Contract

\n

Release Read

This is a paper surface only to the extent the returned artifacts justify it.\nWhen the campaign is partial, the report is still useful because it shows which\nrows returned, which failed, and where the next Distill ribbon should focus. The\npublication claim should be made only from measured pairs.

\n"}