{"id":"internal-science-program","title":"Internal Science Program","subtitle":"The research agenda: what we are trying to learn and how the loop pursues it.","category":"foundations","tags":["program","agenda","vision"],"source":"articles/docs/internal-science-program.md","lang":"en","words":782,"readMinutes":4,"toc":[{"depth":2,"text":"Current Scientific Center","id":"current-scientific-center"},{"depth":2,"text":"Fresh Local Run","id":"fresh-local-run"},{"depth":2,"text":"Claim Triage","id":"claim-triage"},{"depth":3,"text":"Promote","id":"promote"},{"depth":3,"text":"Quarantine","id":"quarantine"},{"depth":3,"text":"Retire Or Rewrite","id":"retire-or-rewrite"},{"depth":2,"text":"Next Internal Experiments","id":"next-internal-experiments"},{"depth":2,"text":"Why This Is Exciting","id":"why-this-is-exciting"}],"html":"

Internal Science Program

This is a private working map for the next evidence push. Do not treat it as a\npublic claim sheet. It is here to keep the research honest while the system\ngrinds.

\n

Current Scientific Center

The strongest idea is not that one potential ranks above another. The strongest\nidea is that interatomic-potential errors can be studied as physical objects:\nthey have covariance spectra, symmetry structure, provenance, failure modes,\nand causal context.

\n

The current center of gravity is:

\n
    \n
  1. Error vectors for elastic observables are strongly compressed.
    2. \n
  2. Benchmark heterogeneity is signal, not nuisance.
    3. \n
  3. Simpson-style claims need strict causal auditing before they are allowed back\ninto headline status.
    4. \n
  4. The next real battlefield is curvature: phonons, Hessians, force constants,\nand non-equilibrium configurations.
    5. \n
  5. Formalization should trail only stable claims with artifacts and falsification\nnotes.
    6. \n
\n

Fresh Local Run

Command run privately into .glim-runtime/science-runs/:

\n
cargo run --release --manifest-path ..\\..\\atlas-distill\\Cargo.toml --bin atlas-distill -- benchmark (Resolve-Path ..\\..\\nist_benchmark.csv) --full\n
\n

Input: nist_benchmark.csv

\n

Dataset summary:

\n\n

Manifold results from the private run:

\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
PotentialMaterialsPR / 5PR CIFirst mode varianceStrict ribbon
Ackland-198741.4581.000-1.69780.5%no
Foiles-198641.6851.000-1.97274.6%no
Zhou-200451.1501.000-1.44293.0%no
Adams-198941.9911.000-1.99165.9%no

Interpretation:

\n\n

Meta-analysis results from the same run:

\n\n

Interpretation:

\n\n

Claim Triage

Promote

Compressed error subspace:

\n\n

Heterogeneity as diagnostic:

\n\n

Curvature validation:

\n\n

Quarantine

Simpson's paradox:

\n\n

Strict hyper-ribbon law:

\n\n

Retire Or Rewrite

Any public wording that says "Simpson's paradox proven" should be treated as\noutdated until the unified causal audit says otherwise.

\n

Any public wording that says "hyper-ribbon proven" should specify which\ncriterion is meant. Low PR is not the same as a strict geometric spectrum.

\n

Next Internal Experiments

    \n

  1. Causal audit matrix

    \n\n
    2. \n

  2. Rank-aware manifold audit

    \n\n
    3. \n

  3. Bulk/shear mode basis

    \n\n
    4. \n

  4. Phonon sentinel protocol

    \n\n
    5. \n

  5. Lean gate preparation

    \n\n
    6. \n
\n

Why This Is Exciting

The scientific path is becoming sharper. GLIM is not merely a validation runner;\nit is a machine for discovering the structure of model error. If that holds,\nthen the useful product is a living error atlas: which observables collapse,\nwhich modes stay stiff, which potential families fail by the same geometry, and\nwhich new experiments break the compression.

\n"}