• Window: 1997-01-01 through the latest published result, using Europe/Prague race dates.
• Grain: one row per horse start in horse_race_starts, keyed by race, horse, and saddle number.
• Cadence: whole-table rebuild; the upstream collection and weather sources do not expose one common bounded incremental cursor.

• Race, entry and connection facts available around race time, including IDs, venue/course, schedule, conditions, advertised prize, horse, stable, weight, equipment, jockey and trainer.
• Observed result facts, including finish/status, time, margin, earnings, equivalent rating, sectionals and race verdict text.
• Retrieval-time horse-profile facts, pedigree and raw profile fields. These describe the current fetched profile and must not be assumed historically point-in-time.
• Weather station mapping and daily observations. Race-day daily aggregates are observed outcomes, not safe pre-race features.
• Raw extraction strings, source URLs, retrieval timestamps, content hashes and a profile-layout validation flag for audit and reparsing.

Column descriptions label fields as PRE-RACE, POST-RACE, RETRIEVAL-TIME, PROVENANCE or QUALITY where leakage risk matters. Targets and rolling ML features are deliberately left for downstream feature engineering.

Absent source values remain null and are never inferred. Original racing statuses, currencies, raw strings, failed pages and unmatched profiles remain distinguishable.

Racing results and profiles come from public Jockey Club Czech Republic pages. Weather observations come from Czech Hydrometeorological Institute open-data files; redistribution rights are not asserted, so the dataset remains private.