Lupine Science is a public research program for understanding where interatomic potentials fail, why those failures have structure, and how that structure can guide correction.

This repository is written first for materials scientists, computational materials groups, MLIP builders, research software teams, and lab leaders who need to decide whether a model can be trusted outside the narrow conditions where it looked accurate.

The central claim is simple but demanding: prediction error is not just noise. Across potentials, elements, properties, and structure families, errors can form low-dimensional geometry. If that geometry is stable, it can tell us what a potential gets wrong, where the next failure is likely, and what correction or new benchmark would actually matter.

Lupine Science is not a finished product page. It is an operating research corpus: claims, evidence, refutations, formalization attempts, viewer artifacts, and a changelog that keeps the self-correction visible.

Modern materials research increasingly depends on interatomic potentials and machine-learned interatomic potentials (MLIPs). The hard question is not only "which model has the lowest benchmark error?" It is:

• Where does this potential fail?
• Are the failures structured or idiosyncratic?
• Does a foundation MLIP inherit the same error geometry as classical models?
• Which apparent trends vanish under matched-sample tests or causal checks?
• Can a claim be inspected, reproduced, refuted, and corrected without losing its provenance?

Lupine Science treats those questions as the scientific object. The goal is to turn model failure into a field of evidence that a human researcher can inspect and an agentic research system can keep extending.

The important cultural point is that refutation is not treated as failure. Self-correction is part of the method. A claim that changes status should become more useful, not disappear.

Start with the public Library:

library.lupine.science

The Library is the human knowledge surface for reports, claim status, evidence summaries, formal notes, and the working changelog. It is generated from this repository, so the corpus is the source of truth and the site is a readable view of it.

Useful local entry points:

LUPI is the browser-native viewer for atomistic evidence:

lupi.live

Use LUPI when the evidence has structures, trajectories, galleries, or visual inspection routes. The viewer is not the whole science; it is the inspectable surface for evidence that benefits from spatial or temporal inspection.

Local code lives under atlas/.

Build the Library locally:

cd library-site
npm install
npm run dev

Run the Rust scientific engine checks:

cargo test --manifest-path atlas-distill/Cargo.toml --bin atlas-distill
cargo clippy --manifest-path atlas-distill/Cargo.toml --bin atlas-distill -- -D warnings

Run the focused repo gates:

just think-lint
just engine-test
just live-build

On Windows, use Git Bash for Node and build tasks. The root justfile already does this with the explicit Git Bash path.

New claims should be written as evidence-bearing research objects, not loose marketing copy.

Use these templates:

Every serious claim should identify the model family, material set, property target, evidence path, status, known confounders, and the next test that could change its status.

Lupine Science is explicit about epistemic status. The exact state changes over time, so treat CHANGELOG.md and the Library as the live record.

For the full root ownership ledger, including keep/elevate/remove-candidate decisions, see ROOTS.md.

The old lupine-start/ marketing/start site, the distiller/ KB, the lupine-distill/ Rust crate, and the lupine-dspy/ package have been retired and archived under archive/. Public research should surface through the Library, the LUPI viewer, and the glim-think feed rather than through a second launch site.

• docs/ONBOARDING.md — research-scientist and software-engineer tracks
• CONTRIBUTING.md — contribution kinds, workflow, verification commands
• scripts/bootstrap.ps1 / scripts/bootstrap.sh — install lightweight dev deps and run quick checks

If you are evaluating the program, the best way to understand it is to watch the public evidence trail rather than look for a pitch surface.

Public-facing surfaces use one naming contract:

Avoid retired organization labels, legacy viewer labels, and retired viewer domains in new copy, metadata, links, and public docs.

Agent-readable files are first-class public artifacts:

After editing canonical agent files, run:

python scripts/sync_brand_agent_text.py

That republishes /llms.txt, /llms-full.txt, and /brand.json into the public static roots.

@unpublished{welcing2026causal,
  author  = {Welcing, Alexander},
  title   = {The Causal Geometry of Prediction Errors in Interatomic Potentials:
             A Hyper-Ribbon Manifold Analysis},
  year    = {2026},
  note    = {Working paper, in preparation}
}

MIT - see LICENSE.

This work builds on sloppy-model theory, causal inference, meta-analysis, materials benchmark infrastructure, OpenKIM/NIST-style potential corpora, and the broader computational materials community.