[
  {
    "section": "s1",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "1. A training run is a data-dependent procedure"
  },
  {
    "section": "s2",
    "kind": "source_based_review",
    "sources": [
      "dl2-a03-1",
      "dl2-a03-2",
      "dl2-a03-3",
      "dl2-a03-4",
      "dl2-a03-5",
      "dl2-a03-6"
    ],
    "claim": "2. The selected papers do not use one interchangeable notion"
  },
  {
    "section": "s3",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "3. Neighboring data and random coupling must be declared"
  },
  {
    "section": "s4",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "4. Loss stability differs from parameter stability"
  },
  {
    "section": "s5",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "5. A quadratic procedure with a closed-form trajectory"
  },
  {
    "section": "s6",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "6. Replacing one observation yields an explicit sensitivity bound"
  },
  {
    "section": "s7",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "7. Expected generalization can also be computed in one special distribution"
  },
  {
    "section": "s8",
    "kind": "source_based_review",
    "sources": [
      "dl2-a03-2",
      "dl2-a03-4"
    ],
    "claim": "8. A bound that worsens with time is not an empirical learning curve"
  },
  {
    "section": "s9",
    "kind": "source_based_review",
    "sources": [
      "dl2-a03-3",
      "dl2-a03-5"
    ],
    "claim": "9. Data-dependent quantities require their own statistical discipline"
  },
  {
    "section": "s10",
    "kind": "source_based_review",
    "sources": [
      "dl2-a03-6"
    ],
    "claim": "10. Expectation and tail guarantees answer different operational questions"
  },
  {
    "section": "s11",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "11. The whole selection pipeline can destabilize a stable learner"
  },
  {
    "section": "s12",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "12. A finite verification artifact with negative controls"
  },
  {
    "section": "s13",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "13. A deep-learning experiment that can test a specific hypothesis"
  },
  {
    "section": "s14",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "14. Stability is useful when its quantifiers remain visible"
  }
]