EmpireCore isn't a model that guesses. It's a seven-stage enrichment pipeline that pre-computes every signal before Claude synthesizes the final pick. This is how institutional-grade intelligence gets built on a $7/month server.
For engineers and investors: When the Fear & Greed Index drops below 25 and VIX spikes above 25, retail participants across all prediction markets — sports, crypto, Polymarket — exhibit correlated irrational behavior. Public bettors over-back favorites by 6.2% above statistical baseline. The 20-40 minute window before prediction markets reprice creates an exploitable edge. No existing platform operationalizes this cross-asset correlation in real time.
In plain English: When the stock market is scary, people bet with their emotions everywhere — sports included. They pile onto favorites, ignore real data, and create opportunities for systematic thinkers. We built a system that detects that fear in real time and bets against the emotional crowd.
Click each gate to see the technical explanation and the plain English version side by side.
Four times a day — 07:00, 12:00, 16:00, 21:00 UTC — the full pipeline runs automatically.
Every day the system fetches game data for all NBA, MLB, and NHL games from ESPN. This includes team matchups, betting lines from multiple sportsbooks, injury reports, and schedule data. The system runs this automatically — no manual input.
Each game passes through all seven intelligence gates above. By the time a game reaches Claude, it has been scored for sharp money divergence, fatigue, consensus divergence, injury impact, player power, archetype classification, and macro regime. This takes approximately 0.5 seconds.
Claude receives a structured briefing with all 7 gate outputs pre-computed. It synthesizes the signals into a final pick with confidence score, Kelly Criterion bet sizing, key signal attribution, and rationale. Claude is the last layer — not the first. It synthesizes pre-computed intelligence, it doesn't generate it from scratch.
Every pick goes through a strict validation pipeline. Spread caps are enforced (NBA max 14.5, MLB/NHL always ±1.5). Matchup reconstruction verifies team names against the real game data. Duplicate suppression prevents the same game being picked twice. Only picks that pass every gate get saved.
After games end, the auto-grader fetches final scores from ESPN and grades every pending pick automatically. No human touches the record. Win, loss, push — all graded against the spread. The full history is public and immutable on the ledger page.
Every graded pick feeds back into the signal performance engine. Signal types that are hitting above baseline get higher weight next harvest. Teams with poor track records get FADE_TARGET flags. The system gets better with every game — automatically, without retraining.
Three or more Tier 1 signals converging. Sharp money confirmed. Macro regime amplifying. Fear window active or archetype discrepancy detected.
One or two strong signals. Line value confirmed. No structural kill switches triggered. Solid play but not maximum conviction.
Single signal, composite only, or marginal edge. Playable but requires smaller sizing. System generates these but recommends caution.
HARD_HOLD from ingress. Away injury >0.60. Spread exceeds physical cap. Missing line data. These picks are never shown — filtered before Claude sees them.
Every pick logged before tip-off. Every result graded automatically. No cherry-picking. No curated samples. The full ledger is live.