Car position encodes as a 1D manifold in mountain car task; linear interventions cross voids and fail; geodesic following enables smooth control.

{
  "doc": "geiger-goodfire-world-inside-neural-networks-2026.md",
  "author": null,
  "context": "Core empirical validation: demonstrates that geometric description correctly predicts intervention outcomes where linear models fail.",
  "enrichment": {
    "is_stale": true
  },
  "norm_label": "Car position encodes as a 1D manifold in mountain car task; linear interventions cross voids and fail; geodesic following enables smooth control.",
  "source_url": null,
  "graphify_id": "finding_mountain_car_1d_manifold",
  "source_file": "geiger-goodfire-world-inside-neural-networks-2026.md",
  "imported_from": "/Users/antonborzov/Documents/Research.nosync/papers/extract_typed_out/geiger-goodfire-world-inside-neural-networks-2026/graph.json",
  "extracted_type": "finding",
  "graphify_file_type": "finding"
}