{"id":"mlip-ni-zero-point-policy-replay","title":"Ni Zero-Point Policy Replay","subtitle":"Local Rust replay turns the Ni promotion canary green before a Cloud Run image rebuild.","category":"validation","tags":["mlip","distill","ni","hyperribbon","live-lab"],"source":"articles/docs/mlip-ni-zero-point-policy-replay.md","lang":"en","words":359,"readMinutes":2,"toc":[{"depth":2,"text":"Why This Matters","id":"why-this-matters"},{"depth":2,"text":"Result","id":"result"},{"depth":2,"text":"Pair Table","id":"pair-table"},{"depth":2,"text":"Interpretation","id":"interpretation"},{"depth":2,"text":"Next Gate","id":"next-gate"}],"html":"

Ni Zero-Point Policy Replay And Canary

Status: local replay gate passed; Cloud Run promotion canary passed
Generated: 2026-05-27
Campaign: ni-fcc-eam-home-turf-paired-accuracy-v1
Replay artifact: library-site/src/reports/assets/mlip/ni-paired-accuracy-zero-point-replay-summary.json
Cloud canary artifact: library-site/src/reports/assets/mlip/ni-paired-accuracy-promotion-canary-summary.json

\n

Why This Matters

The first full Ni paired run rejected our broad transfer ribbon: no measured\npair improved and ten pairs regressed. That was not a launch failure; it was the\nsystem doing its job. The artifacts showed that the prior support family was\nnot appropriate for pure Ni, so the next question became sharper:

\n

Can a same-material, non-overlapping Ni support ribbon improve energy and\nrelaxation accuracy without changing the MLIP or rerunning raw predictions?

\n

The answer from local replay was yes enough to justify a new Cloud Run canary.\nThe rebuilt MACE, CHGNet, and ORB Cloud Run jobs then confirmed the same\ndirection: six improved canary pairs, zero regressions.

\n

Result

\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
MeasureReplayCloud canary
Canary pairs66
Improved pairs66
Regressed pairs00
Mean relative error lift75.12%about 75%
Policy mode appliedmaterial_family_zero_pointmaterial_family_zero_point
Claim statusSpend gate passedAccuracy canary passed

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
RowMLIPBaseline errorCloud Distill errorLift
Energy-volumeMACE-MP-01.28030.320075.01%
Relaxation stabilityMACE-MP-01.27980.318475.12%
Energy-volumeCHGNet1.29430.323675.00%
Relaxation stabilityCHGNet1.29300.319675.28%
Energy-volumeORB-v31.04380.260875.01%
Relaxation stabilityORB-v31.04270.257575.31%

Interpretation

The old policy treated all large energy corrections as dangerous. The Ni canary\nshowed a more specific case: the support set and sealed evaluation set share the\nsame material family, have no structural leakage, pass exact support/eval\ndistance, and show roughly 99.5-99.9% support lift. In that case, the correction\nis best treated as a guarded energy zero-point alignment, not ordinary\ncross-material transfer.

\n

The Rust policy now keeps the old max_energy_bias_ev_per_atom cap for normal\ncorrections and adds a separate\nmax_energy_zero_point_shift_ev_per_atom path that requires:

\n\n

Next Gate

Promote the Ni material-family ribbon to the full 5x5 paired accuracy run. The\npromotion rule stays strict: paired baseline and Distill cells must share raw\nprediction checkpoints, every pair must be measured, and no pair may regress.\nAcceleration remains out of scope until accuracy is locked.

\n"}