{"id":"lupine-refinement-sprint-announcement","title":"Lupine Refinement Sprint Announcement","subtitle":"Five theorems, one benchmark, zero sorry — PR-ready summary of the latest formalization sprint.","category":"validation","tags":["announcement","formalization","press","mlip","benchmark"],"source":"articles/docs/ANNOUNCEMENT-lupine-refinement-sprint-2026-06-27.md","lang":"en","words":812,"readMinutes":4,"toc":[{"depth":2,"text":"TL;DR — what we shipped","id":"tl-dr-what-we-shipped"},{"depth":2,"text":"Why this matters","id":"why-this-matters"},{"depth":2,"text":"The 3×3×3 Layer 2 benchmark campaign","id":"the-3-3-3-layer-2-benchmark-campaign"},{"depth":2,"text":"For the press","id":"for-the-press"},{"depth":3,"text":"One-paragraph summary","id":"one-paragraph-summary"},{"depth":3,"text":"Key talking points","id":"key-talking-points"},{"depth":2,"text":"Social media thread (ready to post)","id":"social-media-thread-ready-to-post"},{"depth":2,"text":"What changed in the repos","id":"what-changed-in-the-repos"},{"depth":2,"text":"Contact / how to follow","id":"contact-how-to-follow"}],"html":"

Lupine Refinement Sprint: Five Theorems, One Benchmark, Zero sorry

Date: 2026-06-27
Announcement type: public / press / investor update
Canonical links:

\n\n
\n

TL;DR — what we shipped

    \n
  1. No-harm guarantee for the 1-D shared-error correction. We proved in Lean that when the shared error lies on a 1-D manifold, the projected correction never increases the residual norm, then aligned the Rust mlip_correct coefficient exactly with the Lean alpha.
    2. \n
  2. Rank-k transferability bound for real alloys. We extended the 1-D sine-of-principal-angle bound to rank-k subspaces and locked the Mg-Li / Al-Cu-style LOOCV reasoning behind a build-locked theorem.
    3. \n
  3. Active sampling acquisition contract. We formalized the greedy residual-max rule and proved both a one-step optimality theorem and a rank-k sample-complexity bound.
    4. \n
  4. Composition optimizer hardening. mlip_optimize now carries elastic-stability guards, moduli-consistency checks, R² refusal, jackknife RMSE, and an optimality-gap estimate.
    5. \n
  5. Epistemic provenance tags. Every benchmark entry is now tagged synthetic, nistIpr, or experiment; the validation report warns when synthetic fixtures are present so they cannot be mistaken for empirical evidence.
    6. \n
\n

All five areas compile: lake build OpenDistillationFactory is green and cargo test / cargo clippy --all-targets -- -D warnings pass.

\n
\n

Why this matters

Machine-learned interatomic potentials (MLIPs) are being asked to predict elastic constants, defect energies, and phase stability for real alloys. Before we trust them for high-stakes materials decisions, three questions need rigorous answers:

\n\n
\n

The 3×3×3 Layer 2 benchmark campaign

We are now running the 56 element/model tasks of the Layer 2 3×3×3 16-element cubic-metal elastic-constant grid concurrently in isolated GCP runs (14 elements × 4 universal MLIPs: M3GNet, CHGNet, TensorNet, QET, each at PBE and r2SCAN). The 4×4×4 supercell comparison is postponed; the 3×3×3 grid gives the precision/cost ratio we need for the public preprint.

\n

The benchmark artifacts are already in the public library:

\n\n
\n

For the press

One-paragraph summary

Lupine, the open MLIP-distillation research platform, has completed a five-front formalization sprint that turns heuristic correction and sampling rules into machine-checked theorems, hardens the composition optimizer against physically impossible recommendations, and tags every benchmark entry with its epistemic source. The work is published in the project’s public library as the team launches a 56-run concurrent GCP benchmark of universal MLIPs across 14 cubic metals.

\n

Key talking points

\n
\n

Social media thread (ready to post)

1/ We just shipped the Lupine refinement sprint: five formalization wins that make MLIP correction, transfer, sampling, optimization, and benchmarking more honest and rigorous.\n\n2/ Theorem 1: a 1-D shared-error correction is *no-harm*. We proved in Lean that the corrected residual norm never exceeds the uncorrected one, then matched the Rust implementation to the proof.\n\n3/ Theorem 2: rank-k alloy transferability. Projecting a target residual onto a source subspace leaves at most sin(θ_k)·‖target‖. Real-alloy LOOCV is now build-locked.\n\n4/ Theorem 3: active sampling has a contract. Greedy residual-max selection is one-step optimal, and the number of informative observations is bounded by the residual subspace rank.\n\n5/ Optimizer hardening: `mlip_optimize` now rejects Born-unstable fits, checks moduli consistency, refuses low R² surfaces, and reports a jackknife RMSE + optimality gap.\n\n6/ Provenance tags: every benchmark entry is now synthetic / nistIpr / experiment, and the validation report warns if synthetic data is present. No more “synthetic fixture reported as empirical.”\n\n7/ We are now running 56 concurrent GCP jobs for the 3×3×3 16-element universal-MLIP elastic-constant benchmark (M3GNet, CHGNet, TensorNet, QET × 14 metals).\n\n8/ All theorems compile with zero `sorry`, and the Rust code passes `cargo test` + `cargo clippy -D warnings`.\n\n9/ Read the public library: https://library.lupine.science\n   Source: https://github.com/alexwelcing/lupine\n\n10/ If you care about trustworthy materials ML, this is the kind of first-principles engineering that has to happen before high-stakes decisions. Thread. 🧵\n
\n
\n

What changed in the repos

\n

All changes are live in main and mirrored to lupine-rhizo and lupine-ledger.

\n
\n

Contact / how to follow

\n"}