Social factors — Phase6 persistence and robustness diagnostics, v3 Study: social-robustness-2026-10-10-v3 10 October2026 This diagnostic phase follows the original eight-phase programme: define tests; build data; test deprivation; test crime; test education/combinations; validate persistence/geographic transfer; publish supported findings; maintain evidence. The exploratory amendment tested all three social blocks. This phase examines persistence, model assumptions, region influence and calibration. It does not complete independent confirmation of the original v1 hypotheses. POST-RESULT STATUS All v2 outcomes, including2025, were already inspected when the diagnostic protocol was written. This is not a new untouched validation set. V1/v2 remain immutable and no hypothesis is promoted by v3. No human actuarial review is recorded. The fixed complete-history cohort uses future knowledge of data availability and remains retrospective. Source-vintage, joint road class/urban-rural and resident-versus-road-location limitations continue to apply. SCOPE Use the same133authority and28primary-police-force cohorts, outcomes, matching miles and paired comparisons as v2. The29force recording-break sensitivity remains in v2. No source updates, imputation or population-based mileage allocations. All eleven original primary-study comparisons are retained. FOUR DIAGNOSTICS 1. Fixed2023parameters: reuse archived2023fit coefficients, scalers, history credibility prior and regional encoding to predict2023–2025. Covariates still update using each year's two-year lags and realised scoring exposure. This tests parameter persistence with updated information, not a three-year forecast issued inJanuary2023. Compare2023reconstruction with original predictions exactly. 2. Poisson: same linear predictors, offsets, training/test rows and ridge1e-6 as original; fit Poisson counts instead ofNB2. This changes large-count weighting as well as variance assumptions. Training Pearson dispersion is retained. A better aggregate score does not validate Poisson uncertainty or establish causation. 3. Fixed2015deprivation: each unit's original2015income/employment context from its2018row is carried through all years. Force scores retain the original fixed2011resident-weighted calculation. Removing the2019switch tests vintage sensitivity; it does not establish the preferred future predictor. 4. No regional intercepts: remove region columns from both candidate and comparator, with all other NB2 features, offsets and folds unchanged. Unseen regions remain excluded from training. Do not choose a replacement winner using inspected outcomes. TRAINING AND ASSESSMENT Refitted variants use the exact v2 primary temporal2023/24/25 expanding folds and2025leave-one-region-out folds. Training excludes2020–2021outcomes. Pandemic lagged predictors remain:2022uses2020history/traffic and2023uses2021. Noholdout calibration correction or parameter selection. Fixed2023is temporal-only; no spatial score is fabricated.504new fits were completed plus reuse of14archived2023models. Convergence diagnostics and transforms are downloadable. REGIONAL INFLUENCE For each originalv2spatial comparison, regional contribution to combined gain equals (regional baseline deviance minus candidate deviance)/total baseline deviance. Contributions sum to combined gain; they are percentage points, not within-region relative gains. Also show each region's mileage share and O/E. The leave-one-region-out-of-scoring column recalculates the score on the remaining original predictions. It does not refit or justify omitting that region from the primary population. FEATURE SUPPORT For each original spatial training fold, find minimum and maximum of every continuous predictor in the expanded model, excluding year and region indicators. Apply the archived training transformations/history prior. Flag test rows outside any marginal range; record all affected features and mileage share. Do not remove or clip flagged rows. Inside all marginal ranges is not proof of joint covariate support. CALIBRATION Use the expanded model's training predicted rates, not counts or assessment outcomes, to form five mileage-weighted risk bands. Apply frozen boundaries to that fold's test rows. Assess candidate and comparator on exactly the same rows within each band. Pooled bands are within-fold ranks, not common numeric rate intervals. Download per-fold cutpoints. Report counts, exposure, region count, observed/predicted rates and O/E. No independent-row confidence intervals are assigned to sparse force bands. METRICS AND UNCERTAINTY Unchanged paired held-out Poisson deviance, gain=1-candidate/baseline deviance, observed/expected and exposure-weighted absolute rate error. A positive gain does not mean fewer collisions or claims. Combined intervals use2000paired resamples of nine region blocks over frozen predictions. They are conditional on fits, exclude refitting/source-measurement uncertainty and are unadjusted across these diagnostic comparisons. Individual-region results are point diagnostics. No new pass/fail promotion threshold is introduced. RESULTS TO INTERPRET The combined authority deprivation/education block retains temporal gain with fixed2023parameters:12.2% versus original11.0%. Its original spatial gain was−4.35%; Poisson gives+6.51%, fixed2015deprivation+3.90%, and removing regional intercepts+7.11%. These sign changes show specification sensitivity; they do not replace the failed original transfer test. London contributes−6.46percentagepoints to the original combined authority spatial gain.24of33London authorities (63.9%ofLondonmatchedmileage) lie outside at least one training predictor's marginal range. OmittingLondonfromscoring would change combined gain to+3.38%, butLondonisretained. Under-transfer/extrapolation must be investigated rather than removed from the study. For burglary beyond deprivation/history, fixed2023parameters still worsen temporal deviance4.11%; Poisson instead improves it6.65%but worsens spatial performance4.67%. Other sensitivities also change the sign or magnitude. Burglary is not a stable general predictor in this study. Income and employment are highly correlated (automatedaudit:0.919authority,0.968force). Individual coefficient directions should not become causal narratives or pricing factors. An automated numerical audit reproduced all294original predictions and found no material v2 calculation error; this is not human actuarial approval. NUMERICAL CORRECTION AUDIT The first v3 attempt stopped because L-BFGS-B's relative-objective stopping left a Poisson absolute gradient0.053246, just above the unchanged0.05gate. The same convex training objective was polished with Newton steps. No features, data, folds, likelihood, ridge, metrics or gate were changed. The failed code/design/log are preserved in attempt-1; numerical-correction.json records the change. No v3comparison summaries were generated or inspected before that correction. Final frozen-design.json identifies the corrected numerical code before the completed diagnostic run. REPRODUCTION Extract reproduce-v3.zip into an empty directory. It contains the exact v2 prerequisites and v3 frozen protocol/design/code. python scripts/social-robustness-v3.py fit Dependencies: Python3, numpy2.3.5, scipy1.17.0 as recorded by the v2 study. The command validates all current frozen input hashes and regenerates results, predictions, fits, band thresholds and sensitivity.csv. Numerical outputs should match; computedAt changes on rerun. Additional regional/calibration CSVs are direct flattened views of results.json. Model code, protocol, failed-attempt audit and hashes are linked on the website. prepare refuses to overwrite an existing frozen design. A new analytical specification requires a new version. NEXT REQUIREMENTS Improve verified road-context controls and source-vintage/population reconciliation; freeze a revised candidate before obtaining fresh outcomes or an independent validation dataset; obtain actual actuarial review. No2025diagnostic can be presented as a fresh unseen confirmation. Recurring model refresh/publication is not enabled by this phase.