{"id":"navigation","title":"Repo Navigation Guide","subtitle":"Quick-search codemap for finding things fast across the workspace.","category":"foundations","tags":["navigation","reference"],"source":"articles/docs/navigation.md","lang":"en","words":1083,"readMinutes":5,"toc":[{"depth":2,"text":"The claim, in one sentence","id":"the-claim-in-one-sentence"},{"depth":2,"text":"The 60-second path to the real info","id":"the-60-second-path-to-the-real-info"},{"depth":2,"text":"The error-geometry objects — disambiguated (read this before \"ribbon\" anything)","id":"the-error-geometry-objects-disambiguated-read-this-before-ribbon-anything"},{"depth":2,"text":"Claim status (live)","id":"claim-status-live"},{"depth":2,"text":"This session's additions (MiniMax-M3 upgrade + the live campaign) — and their status","id":"this-session-s-additions-minimax-m3-upgrade-the-live-campaign-and-their-status"},{"depth":2,"text":"Repo structure and support","id":"repo-structure-and-support"},{"depth":2,"text":"Stale-doc notes","id":"stale-doc-notes"}],"html":"

Start Here — navigating Lupine Science

The fastest path from "what is this?" to the real science. Paths are relative to\nthe repo root (this checkout — there is no glim/ superfolder; an earlier\nversion of this file claimed one, that was stale).

\n

For the narrative front door, read README.md. This file is the\nmap: where the real information lives and the order to read it in.

\n
\n

The claim, in one sentence

Interatomic potentials are necessary and wrong in structured ways: across\nhundreds of potentials their prediction errors collapse onto a low-dimensional\n("hyper-ribbon") structure, and that structure — if it is real and stable — tells\nyou where a model fails and what correction would matter.

\n

The 60-second path to the real info

Read in this order. Each row is the canonical source for that layer.

\n
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
#ReadFor
0docs/ONBOARDING.mdNew contributor? Start here. Research-scientist vs software-engineer tracks, install steps, and common pitfalls.
0.5docs/ARCHITECTURE.mdhow the repo's roots connect into a closed scientific loop
1README.md → "Science Spine"the 7-layer program and why it matters
2archive/swarm_preprint_review/research/immi_dim01_sloppy_theory.mdthe literature foundation — sloppy models, the hyper-ribbon, 25+ primary sources (Transtrum, Waterfall, Frederiksen, Kurniawan…). Start here for the theory.
3lit-review.mdthe assembled review: sloppy theory + Simpson's-paradox/permutation methodology + benchmarking
4docs/data-provenance.mdwhere every number comes from — OpenKIM elastic constants vs NIST, 559 potentials × 15 metals, LAMMPS for Phase-D
5docs/methodology.md · docs/conjectures/ledger.mdhow claims are tested; the live claim ledger (supported / refuted / open)
6paper/immi-paper.texthe IMMI manuscript (the actual paper)
7CHANGELOG.mdwhat changed and what was learned/corrected, newest first
8lean-spec/ · docs/formal-proof-ledger.mdthe formal-specification layer

Deeper theory reports: docs/sloppy_models_report.md,\ndocs/tda_error_landscapes_report.md,\ndocs/phonon_benchmarking_report.md.

\n

The error-geometry objects — disambiguated (read this before "ribbon" anything)

\n

Canonical version: docs/science/objects.md. Read that\nbefore writing "ribbon" anything — it has the full definitions, sources, the\nclaim→object mapping, and the mathematical-coupling caveat. The table below is the\none-screen summary.

\n
\n

Three different objects travel under "low-dimensional error." Conflating them is\nthe single biggest source of confusion in this corpus (the author of this file\nincluded). Keep them straight:

\n
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
ObjectWhere it livesWhat it isCanonical source
A. The sloppy model manifold (the actual hyper-ribbon)data / prediction spaceimage y(θ) of the prediction map as parameters vary; a bounded manifold with a geometric hierarchy of widths Wₙ ~ W₀·ΔⁿTranstrum–Machta–Sethna 2010/2011 — see immi_dim01 §2
B. The empirical effective-dimensionalityobservable space (e.g. C11/C12/C44)participation ratio PR = (Σλ)²/Σλ² of the error covariance across potentials (PR ≈ 1.05–1.86 of 3 → near-1D ribbon). This is how you measure A from data — not a separate manifold.immi_dim01 §4; archive/lupine-distill-rust/src/hypothesis/manifold.rs
C. The configuration-space error coreconfiguration space Ω ⊂ ℝᵐa low-dim core H with the error boundary a codim-1 tube around it; a distinct, more demanding object with a conditional universality theoremthe root PDF "A Conditional Universality Theorem for Error Geometry in MLIPs"

A and B are the established program (B measures A; standard sloppy-model usage). C\nis a separate, rigor-first reframing whose theorem is conditional on nonstandard\nassumptions (notably "A6", that different models share spatial error modes). Bridging\nB→C is not automatic — it needs A6, which had never been tested.

\n

Claim status (live)

The honest record is docs/conjectures/ledger.md and\nCHANGELOG.md. Snapshot from the README: supported —\nclassical hyper-ribbon universality and early de-myopization beyond elastic\nconstants; open / under re-audit — per-element classical→MLIP transfer counts\nafter Born screening, Au escape, Fe magnetic failure mode, predicting E_coh/B0;\nrefuted by us — d-band (sample size), MEAM anomaly (matched-n bootstrap),\nBCC/FCC shield (data contamination).\nPer-conjecture detail: docs/conjectures/.

\n

This session's additions (MiniMax-M3 upgrade + the live campaign) — and their status

Engineering and exploration added 2026-06-02. Read the status column before\ntrusting any of it.

\n
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
ArtifactWhatStatus
docs/glim-m3-upgrade/MiniMax M2.7→M3 model-axis upgrade for the Theorist agent + process docsSolid engineering. The model axis is typechecked, tested. Live M2.7-vs-M3 numbers still need a key.
docs/glim-m3-upgrade/runs/live-campaign-results.mdlive Cloud-Run distill campaign (energy/forces/stress/elastic)Provisional. Real measurements, but it's MLIP energy MAE on MPtrj DFT rows — a different lane from the OpenKIM/NIST elastic-constant corpus the ribbon is built on. The distill "win" is an energy-block recalibration that does not move forces; do not read it as a model improvement.
lean-spec/.../Theory/RibbonProjection.leana kernel-checked parallel/orthogonal correction parabolaToy — mislocated object. It formalizes a scalar decomposition, not the model manifold (A) or the keystone core (C). Keep as a concentration lemma; do not cite as "the ribbon, formalized."
docs/glim-m3-upgrade/runs/a6-alignment-results.mdfirst test of the keystone "A6" shared-mode assumptionProvisional — confound not controlled. Signal is real (same atoms hard across MLIPs) but the test does not yet control for elastic-constant mathematical coupling (Cauchy relation / stability) that Jackson–Somers 1991 and Archie 1981 warn produces a non-zero baseline correlation. Treat as method + first signal only.
docs/science/keystone-reconciliation.mdreconciling the repo with the keystone paperRead with its own correction banner — its original "category error" framing overstated the case (see banner at top of that file).

Repo structure and support

\n

For the machinery (control plane, distill engine, MLIP campaigns, cloud compute,\nthe M3 upgrade), see the engineering index:\ndocs/engineering/README.md. The decision behind this\narrangement: docs/decisions/0002-documentation-architecture.md.

\n

Stale-doc notes

\n"}