Public record
Trust is a system, not a slogan.
Here is the operating method behind our published picks: where inputs come from, when they should refresh, how results are graded, and exactly what a verified badge means.
Methodology
Publish first
A prediction is stored before a completed event can be graded.
Grade mechanically
Completed-game data is compared with the prediction and its captured market line.
Show the full sample
Wins, losses, pushes, and voids remain distinct outcomes; unsupported or missing inputs are voided, not wins.
Source references
supabase/migrations/20260427000010_grade_predictions_multi_sport.sql:25-143
Data sources and sync expectations
| Source | Used for | Harness expectation | Repository evidence |
|---|---|---|---|
| nflverse | NFL schedules, play-by-play, weekly stats, rosters, and injuries | Daily; the health harness allows 25 hours (one hour of drift). | scripts/check-ingest-health.ts:8-10, 26, 81-87 |
| DraftKings DFS | DFS salary snapshots | A successful sync must appear inside the harness’s six-hour freshness budget. | scripts/check-ingest-health.ts:12-14, 27, 149-152 |
| Feed orchestrator | Queued feeds, including odds and Kalshi targets | A completed job must appear inside the harness’s three-hour freshness budget. | scripts/check-ingest-health.ts:16-20, 28 |
Freshness budgets are monitoring thresholds, not a promise that upstream publishers never delay or revise data.
Grading rules
The database grader reads only predictions without an existing result and games with final scores. It grades spread, moneyline, total, supported player-prop, and DFS records; pushes receive zero units and missing or unsupported inputs become voids.
Trigger, honestly described: this is not a per-row SQL
trigger. PostgreSQL pg_cron invokes the idempotent grading function daily at 09:00
UTC, then refreshes reporting views.
Source references
supabase/migrations/20260427000010_grade_predictions_multi_sport.sql:25-143; supabase/migrations/20270811000100_restart_prediction_grading_chain.sql:1-39
Verified badge criteria
A model version earns the badge only when it has a successful benchmark run on an active, non-legacy suite with at least 50 picks. A model card is verified when at least one of its versions qualifies.
The badge means the benchmark met those checks. It does not promise future accuracy, profit, or that every version of the model is verified.
Source: src/lib/registry/verification.ts:3-31
Honesty ledger: what we don’t do
- We do not accept or place bets.
- We do not hold stakes or pay winnings.
- We do not guarantee profit or future performance.
- We do not cherry-pick a model’s record: pushes and voids stay visible as separate outcomes.
Shark Snip is a picks, analytics, simulation, and model-publishing product—not a sportsbook.
Source references
docs/business/faq.md:8-22; supabase/migrations/20260422120000_prediction_results.sql:1-31