[
  {
    "section": "s1",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "1. A teacher is a mechanism, not an extra source of truth"
  },
  {
    "section": "s2",
    "kind": "source_based_review",
    "sources": [
      "dl2-a08-1"
    ],
    "claim": "2. Same capacity does not imply the same training problem"
  },
  {
    "section": "s3",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "3. Soft targets change the gradient direction"
  },
  {
    "section": "s4",
    "kind": "source_based_review",
    "sources": [
      "dl2-a08-2"
    ],
    "claim": "4. Weight averages and output averages are different"
  },
  {
    "section": "s5",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "5. Unrolling a scalar moving teacher"
  },
  {
    "section": "s6",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "6. Stationary noise reduction has strict assumptions"
  },
  {
    "section": "s7",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "7. Smoothing introduces lag under drift"
  },
  {
    "section": "s8",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "8. Consistency admits an uninformative zero-loss map"
  },
  {
    "section": "s9",
    "kind": "source_based_review",
    "sources": [
      "dl2-a08-3",
      "dl2-a08-4"
    ],
    "claim": "9. Successful paired-view methods are complete procedures"
  },
  {
    "section": "s10",
    "kind": "source_based_review",
    "sources": [
      "dl2-a08-5"
    ],
    "claim": "10. Stop-gradient changes dynamics without changing a value"
  },
  {
    "section": "s11",
    "kind": "source_based_review",
    "sources": [
      "dl2-a08-6"
    ],
    "claim": "11. A theoretical mechanism needs its data model"
  },
  {
    "section": "s12",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "12. A finite verification and figure specification"
  },
  {
    "section": "s13",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "13. An empirical study should test the whole learning claim"
  },
  {
    "section": "s14",
    "kind": "author_analysis_or_verification_proposal",
    "sources": [],
    "claim": "14. Better teachers require more than slower movement"
  }
]