[
  {
    "section": "s1",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "1. Interpolation is a constraint, not a complete explanation"
  },
  {
    "section": "s2",
    "kind": "source_based_review",
    "sources": [
      "dl2-a02-1",
      "dl2-a02-2",
      "dl2-a02-3",
      "dl2-a02-4",
      "dl2-a02-5",
      "dl2-a02-6"
    ],
    "claim": "2. The six records study related but nonidentical phenomena"
  },
  {
    "section": "s3",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "3. Fix the estimator and the risk being measured"
  },
  {
    "section": "s4",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "4. A weak singular direction can amplify a label perturbation"
  },
  {
    "section": "s5",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "5. Additional independent coordinates can spread this selected fit"
  },
  {
    "section": "s6",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "6. Duplicating a coordinate can make the norm look better without improving prediction"
  },
  {
    "section": "s7",
    "kind": "source_based_review",
    "sources": [
      "dl2-a02-3"
    ],
    "claim": "7. Effective rank is a family of summaries, not one magic number"
  },
  {
    "section": "s8",
    "kind": "source_based_review",
    "sources": [
      "dl2-a02-4"
    ],
    "claim": "8. The interpolation threshold moves with the algorithm"
  },
  {
    "section": "s9",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "9. A second descent is not guaranteed by a drawing"
  },
  {
    "section": "s10",
    "kind": "source_based_review",
    "sources": [
      "dl2-a02-2",
      "dl2-a02-6"
    ],
    "claim": "10. Noise and signal need separate accounting"
  },
  {
    "section": "s11",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "11. Regularization and finite precision alter the weak directions"
  },
  {
    "section": "s12",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "12. Reproducible finite controls before a neural claim"
  },
  {
    "section": "s13",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "13. An empirical protocol that can challenge the interpretation"
  },
  {
    "section": "s14",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "14. The meaningful claim concerns geometry and procedure"
  }
]