[
  {
    "section": "s1",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "1. Symmetry should preserve the target information"
  },
  {
    "section": "s2",
    "kind": "source_based_review",
    "sources": [
      "dl3-c03-1",
      "dl3-c03-2",
      "dl3-c03-3",
      "dl3-c03-4",
      "dl3-c03-5",
      "dl3-c03-6"
    ],
    "claim": "2. What the selected primary architectures establish"
  },
  {
    "section": "s3",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "3. Proper rotations and reflections are different groups"
  },
  {
    "section": "s4",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "4. Pairwise distances erase orientation"
  },
  {
    "section": "s5",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "5. Signed volume is a minimal handedness witness"
  },
  {
    "section": "s6",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "6. Permutation symmetry must respect the meaning of labels"
  },
  {
    "section": "s7",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "7. Output parity is as important as input equivariance"
  },
  {
    "section": "s8",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "8. A mirror-pair learning lower bound"
  },
  {
    "section": "s9",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "9. Geometric noise creates a separate robustness question"
  },
  {
    "section": "s10",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "10. Energy and biological response require different contracts"
  },
  {
    "section": "s11",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "11. Local neighborhoods can omit relevant relationships"
  },
  {
    "section": "s12",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "12. Dataset splits can reward memorized stereochemical families"
  },
  {
    "section": "s13",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "13. A proposed architecture audit begins with transformation tests"
  },
  {
    "section": "s14",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "14. Differentiation introduces another consistency check"
  },
  {
    "section": "s15",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "15. Handedness must survive the entire prediction pipeline"
  }
]