{
  "id": "social-deciles-validation-2026-10-10-v2",
  "registeredAt": "2026-10-10T17:36:01.730619+00:00",
  "baseStudy": "social-deciles-2026-10-10-v1",
  "amendment": "Phase 5 operational details frozen after v1 temporal point estimates were inspected, before this extension is fitted. Original decision thresholds unchanged. Entire study exploratory.",
  "primaryAndSecondaryComparisons": [
    [
      "combined",
      "DE",
      "B"
    ],
    [
      "deprivation",
      "D",
      "B"
    ],
    [
      "education",
      "E",
      "B"
    ],
    [
      "education-added-last",
      "DE",
      "D"
    ],
    [
      "deprivation-added-last",
      "DE",
      "E"
    ],
    [
      "deciles-vs-linear",
      "DE",
      "C-linear"
    ],
    [
      "deciles-vs-smooth",
      "DE",
      "C-smooth"
    ]
  ],
  "gates": {
    "candidate": "DE vs B gain >=1% pooled, refitted paired spatial-block 95% lower >0; >=1% in each assessment year; O/E 0.90\u20131.10 pooled and each year; positive total leave-region-out gain and >=5/9 positive regions; overall absolute rate error no worse than B; all support gates met",
    "nonInferiority": "DE vs prespecified C-smooth: upper 95% relative excess deviance <0.5%, with no deterioration in overall calibration distance |O/E-1| or absolute rate error",
    "superiority": "DE vs C-smooth gain >=1% and refitted paired-block 95% lower >0; linear comparison secondary",
    "validation": "Nine leave-region-out refits and paired spatial-block refit uncertainty remain Phase 5; no pass/fail promotion at Phase 4. Record access to future outcomes before allocating a fresh holdout."
  },
  "spatial": "Exactly the nine frozen futureSpatialFolds: held region absent from all training years; assess 2025. Training-only cuts, knots, scaling, history prior and unseen-region intercept.",
  "uncertainty": {
    "seed": 202610105,
    "replicates": 999,
    "blocks": "Nine English regions; draw nine with replacement for each replicate. Use the same region multiplicities across every year, fold and model, retaining all years of each area together.",
    "refitting": "Refit all six models for each of the three expanding temporal folds in every replicate, including all preprocessing. Sampled cluster copies receive replica area IDs so each sampled area has one equal-weight record per feature vintage; repeated years never weight cutpoints. Original IDs retained for support counts.",
    "assessment": "Evaluate on the paired assessment-year rows from those same sampled regions, with their multiplicities. No random row splits, outcome leakage or frozen-prediction-only resampling.",
    "interval": "Two-sided percentile 2.5% and 97.5% quantiles, linear interpolation. Gain = 1 - candidate deviance / comparator deviance. Excess = -gain. Limited independence: only nine regional blocks.",
    "failures": "No retries with changed estimators or replacement draws. Archive every fit failure. Require all 999 paired replicates for a gate-eligible interval; intervals with missing fits are diagnostic only. A finite full-rank point design is required. Bootstrap rank/support failures are counted and disclosed, not silently filtered.",
    "rateIntervals": "Refitted paired-block intervals for mean held-out predicted rates within fixed v1 display bands. These include sample composition and parameter uncertainty, condition on realised mileage, and omit exposure measurement error. Not individual predictive intervals."
  },
  "support": "Every original temporal and spatial training feature/vintage/band must contain >=10 distinct original authorities and >=100 collisions; empty bands fail. Bootstrap duplicated areas do not inflate support. No quintile rescue.",
  "sensitivities": {
    "fitted": [
      "poisson",
      "fixed-imd2015",
      "education-level4-complement",
      "include-pandemic",
      "fixed-2023"
    ],
    "influence": "Remove each region from both training and assessment, refit all six models on all three temporal folds. Report every leave-one-region-out result, including London.",
    "poisson": "Same predictors and fixed numerical ridge; inherited convex Poisson optimiser with training-only Newton polishing.",
    "fixed-imd2015": "Use each authority\u2019s observed 2015 income/employment scores in all years; set vintage to 2015.",
    "education-level4-complement": "Replace no-qualifications share with 1 minus the 2011 Level 4+ share; high scores still mean lower educational attainment. Separate alternative, not equivalent qualifications.",
    "include-pandemic": "Include 2020/2021 outcome years in training only where earlier than assessment; primary still excludes them.",
    "fixed-2023": "Freeze the 2023 training rows, coefficients, preprocessing and cutpoints for all 2023\u20132025 assessments. Lagged inputs and realised miles still update.",
    "pandemicStress": "Separate 2020 and 2021 assessment using only 2018/2019 training; not part of the persistence gate.",
    "boundaryChoices": "Unavailable: no audited longitudinal reallocation with matched collision/mileage denominators. Exact-code cohort limitation retained.",
    "ruralUrbanInteraction": "Unavailable: no matched historical road-class \u00d7 urban/rural exposure. 2021 rurality published in 2026 and 2026 network snapshots are not historical controls."
  },
  "classification": "All five baseline/block comparisons use their own candidate/calibration/transfer/error/support rules. If nonpositive pooled gain: no incremental gain. If fits/precision/support unavailable: insufficient evidence. Otherwise failure of any criterion: inconsistent. Smooth/linear comparisons additionally show non-inferiority and superiority; replacement requires DE candidate gate first. Secondary comparisons are exploratory and not multiplicity adjusted.",
  "publication": "One findings.json drives Data Lab and Lens. Original v1 remains pinned. No pricing deployment or human review is implied."
}