{
  "schemaVersion": "data-lab-study/1",
  "id": "london-qualifications-reproducible-v1",
  "title": "London regional baseline and qualifications",
  "question": "Does no-qualifications context add predictive information after choosing a London-specific regional baseline, and how sensitive is the finding to mileage assumptions?",
  "intendedUse": "Exploratory, reproducible actuarial research; not an individual driver score or production pricing factor.",
  "outcome": "Police-reported personal-injury collisions at road location per 100000 all-motor vehicle miles",
  "exposureDefinition": "Matched authority-year DfT all-motor vehicle miles on all roads; rate = 100000 × SUM(collisions) / SUM(miles). Includes non-motor-only injury collisions in numerator.",
  "population": {
    "nationalAuthorities": 133,
    "londonAuthorities": 33,
    "inputRows": 1064,
    "inputYears": "2018–2025",
    "assessmentYears": [
      2023,
      2024,
      2025
    ],
    "assessmentRows": 99,
    "trainingExclusions": [
      2020,
      2021
    ]
  },
  "comparisons": {
    "L": [
      "L",
      "N"
    ],
    "P": [
      "P",
      "N"
    ],
    "S": [
      "S",
      "N"
    ],
    "oldL": [
      "oldL",
      "N"
    ],
    "E-linear": [
      "E-linear",
      "S"
    ],
    "D-linear": [
      "D-linear",
      "S"
    ],
    "DE-linear": [
      "DE-linear",
      "S"
    ],
    "E-bands": [
      "E-bands",
      "S"
    ],
    "D-bands": [
      "D-bands",
      "S"
    ],
    "DE-bands": [
      "DE-bands",
      "S"
    ],
    "E-added-last-linear": [
      "DE-linear",
      "D-linear"
    ],
    "D-added-last-linear": [
      "DE-linear",
      "E-linear"
    ],
    "E-added-last-bands": [
      "DE-bands",
      "D-bands"
    ],
    "D-added-last-bands": [
      "DE-bands",
      "E-bands"
    ],
    "K": [
      "K",
      "N"
    ],
    "S1": [
      "S1",
      "N"
    ]
  },
  "sourceManifest": "source-manifest.json",
  "dataDictionary": "data-dictionary.json",
  "assumptions": "assumptions.json",
  "executableAnalysis": {
    "entrypoint": "reproduce.py",
    "dependencies": "requirements.txt",
    "originalProtocol": "original/protocol.json",
    "amendment": "original/expanded-protocol.json",
    "sensitivityProtocol": "sensitivity-protocol.json"
  },
  "referenceOutputs": {
    "metrics": "original/metrics.csv",
    "predictions": "original/predictions.csv",
    "totals": {
      "assessmentRows": 99,
      "collisions": 64815,
      "motorMiles": 58278600000.0,
      "ratePer100k": 0.11121578074970916
    },
    "models": "original/expanded/point.json.gz",
    "uncertainty": "original/expanded/uncertainty.json"
  },
  "validation": {
    "record": "outputs/validation.json",
    "arithmeticTolerance": {
      "absolute": 1e-08,
      "relative": 1e-09
    },
    "refitPredictionTolerance": {
      "absolute": 1e-06,
      "relative": 1e-08
    },
    "requirements": [
      "Exact row keys and counts",
      "All original bytes match hashes",
      "Fresh point, borough-holdout and central-pair predictions reproduce within tolerance",
      "All five mileage scenarios complete and reconcile",
      "Bootstrap summary matches reference; explicit full-refit status",
      "No model promotion from numerical verification"
    ]
  },
  "interpretation": "The regional baseline is an exploratory candidate. Qualifications alone as a linear term worsen all annual assessments. Bands have insufficient support and uncertainty crossing zero. No social block passes every prespecified check.",
  "limitations": [
    "All assessment outcomes were previously inspected.",
    "Only 33 London authorities; no LSOA exposure or individual driver inference.",
    "Central-borough exclusion materially changes the result.",
    "Mileage stress bounds are assumptions, not measured confidence limits.",
    "Reproduction starts at the frozen model-input snapshot, not raw publisher extraction.",
    "Human actuarial validation is not recorded."
  ],
  "inspiration": "Claude Code for Actuaries — The Complete Guide, E. C. Nzati and D. Schraub (2026): executable study folders and specification-led checks. The paper is not redistributed."
}