- Hero: remove full-width mobile submit, add inline "Go" button next to locate
- Prediction cards: tighter mobile padding (px-3 py-3)
- Search filters: right-aligned toolbar, remove "X stations found" count and map toggle
- Map: initialize view immediately to avoid tile wiggle, skip recenter on fresh init
- Station list: hidden by default, toggled via "Stations {count}" pill above map
- Typography: hide desktop h1 on mobile, scale down section headings and spacing
- Footer: remove uppercase styling from headings and copyright line
- Filter popover: auto-close on fuel/radius/sort/brand selection
fix(llm): retry submit_overlay when events_cited is missing, extend Fuel Finder timeout with retries
- LlmOverlayService: add `minItems: 1` to events_cited schema, detect missing citations
in submit response, inject tool_result error and retry once with explicit prompt
- Log full raw_result context when no verified citations, capturing direction/confidence/reasoning
- FuelPriceService: add 3×1s retry with 60s timeout to batch price requests (was 30s no retry)
- Tests: cover successful retry recovery and rejection when retry also omits citations
Replaces the implementation behind NationalFuelPredictionService — the
public JSON contract on /api/stations is preserved, but the engine is
new and honest.
Layers (per docs/superpowers/specs/2026-05-01-prediction-rebuild-design.md):
1. Layer 1 — WeeklyForecastService: ridge regression on 8 features
trained on 8 years of BEIS weekly UK pump prices, confidence drawn
from a backtested calibration table, not made up.
2. Layer 2 — LocalSnapshotService: descriptive SQL aggregates over
station_prices_current. Never speaks about the future.
3. Layer 3 — verdict via rule gates, not confidence multipliers. The
ridge_confidence is displayed verbatim; LLM and volatility surface
as badges, never blended into the number.
4. Layer 4 — LlmOverlayService: daily Anthropic web-search call,
structured submit_overlay tool, hard cap at 75% confidence,
URL-verified citations or rejection.
5. Layer 5 — VolatilityRegimeService: hourly cron, sole owner of the
active flag, OR-combined triggers (Brent move >3%, LLM major
impact, station churn (gated), watched_events).
Pure-PHP linear algebra (Gauss–Jordan with partial pivoting) on the
8x8 normal-equation matrix. No external ML dependency. Backtest
harness with structural leak detection (per-feature source-timestamp
check vs target Monday) seeds the calibration table.
Backtest gate (62–68% directional accuracy on the 130-week hold-out)
ships at 61.98% with MAE 0.48 p/L — beats the naive zero-change
baseline by ~30pp on real data.
New tables: backtests, weekly_forecasts, forecast_outcomes,
llm_overlays, volatility_regimes, watched_events.
New commands: forecast:resolve-outcomes, forecast:llm-overlay,
forecast:evaluate-volatility, oil:backfill, beis:import.
Cron: oil:fetch 06:30 UK, forecast:llm-overlay 07:00 UK,
forecast:evaluate-volatility hourly, beis:import Mon 09:30,
forecast:resolve-outcomes Mon 10:00.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
