{"id":"mlip-mptrj-broad-dft-canary","title":"MPtrj Broad-DFT MLIP Canary","subtitle":"Support-floor v2 improves MACE, ORB, and SevenNet while CHGNet safely holds: six wins, two holds, zero regressions.","category":"validation","tags":["mlip","distill","mptrj","dft","cloud-run","live-lab"],"source":"articles/docs/mlip-mptrj-broad-dft-canary.md","lang":"en","words":453,"readMinutes":2,"toc":[{"depth":2,"text":"Why This Matters","id":"why-this-matters"},{"depth":2,"text":"Cloud Result: Support-Floor v2","id":"cloud-result-support-floor-v2"},{"depth":2,"text":"Local v3 Replay","id":"local-v3-replay"},{"depth":2,"text":"Interpretation","id":"interpretation"},{"depth":2,"text":"Evidence Contract","id":"evidence-contract"},{"depth":2,"text":"Next Gate","id":"next-gate"}],"html":"

MPtrj Broad-DFT MLIP Promotion Canary

\n

Why This Matters

Nickel is a controlled first lane, not the limit of the Distill claim. This\ncanary asks whether the same paired evidence contract can show improvement on a\nnon-Ni, broad DFT trajectory fixture from MPtrj across several MLIP backends.

\n

The first global ribbon improved MACE, ORB, and SevenNet but regressed CHGNet.\nThe support-floor v2 policy fixed that failure without a CHGNet exception:\nCHGNet now refuses correction because its support residual is already below the\nphysical floor. That gives us a clean cloud result: six improved pairs, two\nsafe holds, zero regressions.

\n

Cloud Result: Support-Floor v2

\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
Energy-volumeMACE-MP-00.4116 eV/atom MAE0.2038 eV/atom MAE50.49%Improved
Energy-volumeCHGNet0.1035 eV/atom MAE0.1035 eV/atom MAE0.00%Safe hold
Energy-volumeORB-v30.4295 eV/atom MAE0.4233 eV/atom MAE1.44%Improved
Energy-volumeSevenNet0.3997 eV/atom MAE0.2795 eV/atom MAE30.06%Improved
Relaxation stabilityMACE-MP-00.5604 penalty0.3866 penalty31.02%Improved
Relaxation stabilityCHGNet0.0557 penalty0.0557 penalty0.00%Safe hold
Relaxation stabilityORB-v30.5327 penalty0.3365 penalty36.83%Improved
Relaxation stabilitySevenNet0.5750 penalty0.3972 penalty30.92%Improved

Summary: 16 / 16 cells completed, 8 paired comparisons measured, 6 improved,\n2 unchanged, 0 regressed. The gate verdict is\npromotable_accuracy_candidate.

\n

Local v3 Replay

The next local iteration keeps the v2 energy behavior, then adds a\nrelaxation_stability row override. It is intentionally not the default cloud\npolicy yet.

\n

Replay result: 8 / 8 pairs measured, 6 improved, 2 unchanged, 0 regressed.\nMean pair lift improves from about 22.6% in v2 replay to 25.6% in v3 replay.

\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
RowMLIPv2 Distillv3 replayv3 lift
Relaxation stabilityMACE-MP-00.38660.328641.36%
Relaxation stabilityORB-v30.33650.320039.94%
Relaxation stabilitySevenNet0.39720.338041.23%

Interpretation

This is the growth loop working correctly. The v1 failure became a v2 refusal\nguard. Then local replay found that energy and relaxation want different policy\nsettings. That row-aware split is exactly the kind of product behavior\nresearchers need: Distill changes the outcome only where the evidence says the\nruntime is inside the support tube, and it refuses an already-strong baseline\ninstead of forcing a correction.

\n

Evidence Contract

\n

Next Gate

Run the four-MLIP MPtrj Cloud Run canary with\nhyperribbon-mptrj-row-hybrid-v3-accuracy. It must preserve the same rule:\nall paired cells complete, every baseline/Distill pair shares a raw prediction\ncheckpoint, no pair may regress, CHGNet must remain unchanged or improve, and\nacceleration remains out of scope until accuracy is locked.

\n"}