Public docs center: `/docs/intro` on the main site (bilingual product documentation, including intraday analysis, calibrated probability, model stack, TAF, settlement sources, history, and extension).
[](https://star-history.com/#yangyuan-zhen/PolyWeather&Date)
- Runtime state, cache, and core offline training/backfill flows now use SQLite as the primary path; legacy JSON/JSONL files remain only for migration, export, and explicit fallback input.
- EMOS/CRPS calibration is wired and trainable, but production should stay on `legacy` or `emos_shadow`; `emos_primary` is only for candidates that pass local offline evaluation and manual rollout.
- Intraday analysis is now positioned as a professional meteorology read: headline, confidence, base/upside/downside paths, next observation point, evidence chain, failure modes, and confirmation rules.
- Intraday modal now blocks stale cached detail during refresh, so users do not briefly trade off old city/date data before full detail arrives.
- City decision cards now include the AI airport read: METAR, DEB, model cluster, and the AI expected-high center are resolved before mapping the result to Polymarket temperature buckets.
- AI airport reads now use in-page memory cache, browser `localStorage`, and backend short-TTL cache; returning from another dashboard tab restores existing stream text or final results before any new request is needed.
- Market bucket matching now uses the full `all_buckets` surface and strict exact / range / or-higher / or-lower direction checks, reducing bad matches to unreasonable tail buckets.
- The card label “model-market difference” means `model probability - market-implied probability`; positive values indicate weather probability above market pricing, while negative values indicate the YES is already priced more fully.
- Calibrated model probability is now the primary probability panel. It shows the active production probability engine; EMOS/LGBM are surfaced only when evaluated or shadowed, while model consensus and market prices remain secondary references.
- Frontend design system overhauled: unified CSS token system, eliminated `!important` abuse (68→6 in light theme), consolidated breakpoints (18→10), migrated hardcoded colors to CSS variables, added ARIA attributes and focus-visible keyboard navigation. See `docs/frontend-ui-design-review.md` for the full audit trail.
- Public in repo: weather aggregation, core analysis, dashboard, bot baseline, and standard payment flow.
- Not included in this repository: private production data, internal operating thresholds, commercial risk rules, pricing strategy details, and growth tooling.
- Trademark, brand, domain, production databases, and hosted-service operations are **not** granted by the code license.
- Generates settlement-oriented calibrated probability buckets (`mu` + bucket distribution), with `LGBM` metadata surfaced when the calibrated engine is active.
- Adds official nearby-network and runway-level enhancement layers for China, Japan, Korea (AMOS runway sensors for Seoul/Busan), Hong Kong, Taiwan, and Turkey without replacing airport settlement anchors.
- APAC: Seoul, Busan, Hong Kong, Lau Fau Shan, Taipei, Shanghai, Beijing, Qingdao, Wuhan, Chengdu, Chongqing, Shenzhen, Guangzhou, Singapore, Tokyo, Kuala Lumpur, Jakarta, Manila, Wellington
- Americas: Toronto, New York, Los Angeles, San Francisco, Aurora, Austin, Houston, Chicago, Dallas, Miami, Atlanta, Seattle, Mexico City, Buenos Aires, Sao Paulo, Panama City
- Airport-linked contracts use the METAR / airport primary observing site as the settlement anchor. Wunderground pages are reference/history pages, not stations.
- Taipei and Shenzhen retain their explicitly configured station history pages for reconciliation, but the docs avoid describing Wunderground itself as a physical station.
Do not run full EMOS retraining on a small VPS. The VPS should collect data and load approved calibration files; training should run on a local/dev machine using a copied production SQLite database:
Promote a generated `default.json` only when `auto_retrain_report.json` has `ready_for_promotion=true`, and prefer `emos_shadow` before enabling `emos_primary`.
