Bucket Blitz
Who is filling it up tonight?
There are no picks on this page, and there will not be until the work below is finished. Not built. The NBA surface is a settled archive — the league is out of season and no live board or player model is running.
What exists today
- Games captured100 confirmed 2026-27 dates — the schedule is captured as the league releases it, never inferred
- Season opensOctober 3, 2026
What has to be true before it publishes
4 of 10 stages complete. Every sport on this site passes the same list — the two that publish today did, and this one has not yet. A stage only counts as done when there is a receipt for it.
- Schedule/event discoveryDone
Counts as done when: daily schedule artifact from an approved source, date-correct in ET.
- Team/player identity + assetsDone
Counts as done when: canonical ids resolving through the identity components with fallbacks.
- Pick qualification policyDone
Counts as done when: a written policy with thresholds, applied by code, with no-play as a first-class outcome.
- Product generationDone
Counts as done when: boards/cards generated by scheduled automation with the product-state contract.
- Required stats/context dataPart-way
Counts as done when: the model's input fields present for a full slate, no paid call outside budget.
- Market/odds ingestionPart-way
Counts as done when: credit-guarded ingestion producing de-vigged implied probabilities.
- Model/simulationPart-way
Counts as done when: a versioned model producing per-event outputs from pregame inputs only.
- Calibration/validationBlocked outside our control
Counts as done when: a preregistered backtest against settled results — the MLB bar (Brier vs market), not a smaller one.
- Publication + freshnessPart-way
Counts as done when: public surfaces derive freshness from the product's own artifact.
- Official settlementPart-way
Counts as done when: grading from official results with lineage stamps, quarantine on ambiguity.
Why it is not shipped early
A model that has never been checked against settled results can still produce confident-looking numbers — that is the easy part. Four models on this site were built, measured against a bar set in advance, and rejected for failing it; their write-ups are public. Bucket Blitz gets the same treatment, which means it appears here when it earns a page, and not before.