Acquire
Checksum archives, verify ZIP CRCs, and record source coverage privately.
Methods and limitations
Lineage before confidence.
The active model is whichever evaluated option clears every gate. Otherwise, the seasonal-naive baseline remains in service.
Preserve
Load each analytical source row exactly once into immutable raw tables.
Contract
Build source-native tables separately from derived forecasting marts.
Evaluate
Run at least twelve rolling origins and gate challengers against seasonal naive.
Release
Export immutable artifacts; promote only by replacing current.json.
MODEL GATES
Promotion is earned.
- At least 5% WAPE improvement at every supported horizon
- MASE below 1
- 75%–90% empirical coverage for the nominal 80% interval
- No greater than 10% WAPE regression in 90% of eligible sub-ICBs
HARD LIMITS
What is not claimed.
- No operational slots, rosters, cancellations, or actual capacity
- No combined “total demand” across overlapping channels
- No deprivation or patient-experience measures as forecast inputs
- No NHS endorsement, clinical guidance, or commercial licence
LINEAGE
From publisher row to bounded release.
Private archives
SHA-256 · CRCPostgreSQL raw
one immutable row keydbt contracts
source-native then derivedModel gates
12+ rolling originsR2 release
atomic current.jsonWorker API
bounded, no DB credentials
SHA-256 · CRCPostgreSQL raw
one immutable row keydbt contracts
source-native then derivedModel gates
12+ rolling originsR2 release
atomic current.jsonWorker API
bounded, no DB credentials
BASELINE AND CHALLENGERS
The baseline earned the release.
| Model | Role | Quantiles | Full-snapshot result |
|---|---|---|---|
| Same-weekday seasonal naive | Required baseline and fallback | p10 · p50 · p90 | Approved · 10.96% WAPE at 28 days |
| Elastic Net | Linear benchmark | Residual interval | Failed 7-day gate |
| LightGBM | Global direct-horizon challenger | Quantile objectives | Failed 7-day gate |
| CatBoost | Global direct-horizon challenger | Quantile objectives | Failed 7-day gate |
Twelve rolling origins covered 104 eligible sub-ICBs. The baseline interval captured 55.09% of observations at 28 days, below its nominal 80%; treat the band as indicative, not a calibrated probability guarantee.
ARCHITECTURE EVIDENCE
Public runtime stays small.
Typed Python and dbt run privately. R2 holds immutable release objects. The Worker streams bounded pages and scenarios; it receives no PostgreSQL credential.
PORTFOLIO EVIDENCE
Claims remain inspectable.
Decision records, source and metric contracts, evaluation status, rollback design, generated-fixture tests, Worker dry-run, responsive browser checks, and WCAG evidence are versioned with the project.