[MaxGhenis]

Adds the v2b distributional deviation model on top of the corrected pipeline (#7): a full per-case predictive distribution of D = RAWBEN − BENFIX, exportable for in-browser Monte Carlo.

Components

• Sign model P(D>0 | deviates), nested out-of-fold isotonic calibration (FY2024 AUC 0.700)
• Conditional quantiles of log|D| at 9 levels, per-case monotone; Pareto tail beyond q99 (fitted scale 0.467, finite-mean condition holds)
• Draw machinery (analysis/predictive_process.py), unit-tested against analytic expectations
• Self-contained app export app/public/model_data.json (44,800 CASE==1 records, 53 jurisdictions; 783KB gzipped)

Validation (FY2024, weighted)

• Quantile coverage within 3pp at 7 of 9 levels (q.75 −3.5pp, q.90 −3.3pp flagged)
• Distributional vs direct stage-2 crossing rates: national 13.22% vs 12.92% (observed 13.39%); equal-state MAE 3.08 vs 3.02pp raw, 1.59 vs 1.56pp with FY2023-fit frozen factors
• Model-vs-bootstrap simulated measured-rate table: the model's raw level underpredicts high-error states (NY −3.1pp) — documented in FINDINGS; app integration anchors levels on the model's own baseline point
• Two byte-identical regenerations, independently reproduced (SHA256 match on all four artifacts); 46 tests; ruff clean

Simulator integration is intentionally not in this PR — app wiring follows separately.

🤖 Generated with Claude Code