ADJACENT DECILE FUSION V3

Exploratory authority-level research. No individual rating factor or LSOA coefficient is deployed.

The point analysis comprises three temporal and nine leave-region-out assessments. Seven fusion penalty ratios are selected inside each training set using three regional inner blocks. Support is assessed using distinct original authorities. Every assessment uses the same collision counts and motor mileage for all six models.

The complete final bootstrap archive contains 199 draws with all preprocessing, tuning and models refitted. Read findings.json for completion, uncertainty and rank diagnostics. Compare point predictions with predictions.csv; point-fits.json.xz contains every training transform and selected coefficient.

Reproduce with the instructions in reproduce-v3.zip. Dependencies: Python 3.12, NumPy 2.3.5 and SciPy 1.17.0. No downloads are required. Set OPENBLAS_NUM_THREADS=1 before running.

The exact all-fused endpoint was corrected to avoid optimizer-roundoff differences being labelled as separate bands; see numerical-amendment.json and the preserved first-attempt point files. checkpoint-audit.json records the subsequent complete rerun and source locations for original-attempt bootstrap archives. The final bootstrap archive is public.

Original decile studies v1 and v2 remain unchanged. Existing prospective 2027–2029 reserve rules remain in force. Browser and mobile visual verification were unavailable in this environment; calculation, render, query-validation and build checks are separate from visual QA.