Investigations
Investigations are closed methodological inquiries — distinct from Methodology pages, which describe how the current production model works. Each investigation poses a specific question, defines a verdict criterion in advance where possible, gathers evidence, and reaches a conclusion that is either adopted into the pipeline or parked with an explicit revisit condition. Promoting an investigation here means the evidence is complete, the verdict is defensible, and the finding is stable enough to describe publicly.
Not all closed investigations are promoted here. Some remain internal diagnostics or are incomplete. The full internal record, including entries for EB shrinkage, facility-family split, and OSM tiered imputation, is in docs/internal/decision-register.md.
| Investigation | Verdict | Summary |
|---|---|---|
| KSI atlas diagnostic | Parked | Severity-reporting inconsistency across police forces makes a standalone KSI model indefensible at current scope. |
| Staffordshire data quality discontinuity | Source-data issue confirmed | Persistent Staffordshire flags are a DfT-acknowledged under-reporting issue, not a pipeline defect; Staffordshire treated as out-of-scope by default for KSI revisit. |
| Temporal descriptors evaluation | Below threshold | Overnight-ratio and HGV descriptors carry real but marginal signal; both failed the pre-registered adoption rule. |
| AADF counted-only filter | Adopted | Filtering to directly-counted AADF rows raised Stage 1a CV R² from 0.72 to 0.83; Estimated-weighted alternative rejected. |
| Rank stability harness | Noise floor established | Five-seed evaluation sets the baseline for evaluating future feature additions. |
| Zero-calibration diagnostic | NB warranted; ZINB not required | Poisson fails the zero-calibration check (p = 0.000); NB with α = 2.057 reproduces the observed zero count (p = 0.722), confirming overdispersion — not zero-inflation — as the dominant distributional feature. |