The Brain

The handicapper is Claude Opus 4.7 — Anthropic's most capable reasoning model — driven by a system prompt that codifies sharp Vegas discipline (no hype, no fire emojis, no "lock" language, edge-required, max -150 juice). For the 5 PM late-edge check, a cheaper Claude Sonnet 4.6 pass runs a focused scan for material new info.

The Data

Per game, the engine harvests:

• Probable pitchers + season + last-3-start splits (ERA, FIP, xFIP, SIERA, K%, BB%, HR/9, Stuff+, WHIP) via the MLB Stats API
• Team-level offense splits — wRC+ vs LHP/RHP, recent 14-day runs/game
• Bullpen quality — rolling 14-day FIP, closer availability
• Statcast park factors for runs and HR (3-year rolling) across all 30 MLB parks
• Weather — temp, wind direction + speed at outdoor venues, modeled HR impact (±15% from wind, ±5% from temp)
• NBA team ratings — Net/Off/Def Rating, eFG%, TS%, pace, rest days
• NHL standings + goalie matchup data
• Live odds across DraftKings, FanDuel, BetMGM with line movement tracked since open

Pre-Computed Signals

Before the model sees the data, the engine pre-computes signal tags it can latch onto — saving tokens and forcing structured reasoning. Signals include:

• Reverse line movement — line drifting against the popular side, a known sharp-money tell
• Pitcher form divergence — last-3 xFIP vs season xFIP gap ≥ 0.6
• Lineup handedness mismatches — wRC+ gap ≥ 15 between team and opposing handedness
• Bullpen disparity — 14-day FIP gap ≥ 1.5 runs
• Weather HR impact ≥ ±8%
• Park-factor extremes — runs PF ≥ 1.07 or ≤ 0.94

Real-Time Web Search

Inside the model call, Scott Bot has tool access to live web search. He uses it for late-breaking signal — confirmed lineup news, surprise scratches, late weather changes, breaking injuries. He doesn't use it for things already in the IntelPack, and he doesn't ramble. Targeted queries only.

The Validator

After Scott Bot returns picks, 10 deterministic rules gate every selection before it ships. The validator isn't Claude — it's pure logic, and if a pick fails, it doesn't make the site.

• Max -150 juice on straight bets
• Confidence-to-units must match exactly (5 → 2.0u, 4 → 1.5u, 3 → 1.0u)
• Pick's data confidence ≥ 0.6
• Only ML, spread, total, prop, or 2-leg parlay — no teasers, SGPs, live bets, or correlated plays
• No same-game opposite sides
• MLB-specific: HR props need HR/9 + L3 ERA data; run lines need top-10 R/G; unders downgrade on starters with BB% > 11
• Ladder pick odds must land between -125 and +130 (~even money)
• Parlay legs must be uncorrelated (different games)

The Ladder Challenge

One pick per day is the ladder pick — priced near even money so a win doubles the unit. The target is 10 wins in a row; pushes carry the streak, losses reset it. When the strongest single play on the board is too short (worse than -125), Scott Bot can build a 2-leg parlay from uncorrelated games to land near even money. The ladder is about discipline, not heat.

The Schedule

• 11 AM ET — Morning Lock. Canonical picks for the day. Once locked, they never change.
• 1 PM ET — Midday Refresh. Re-snapshot odds for CLV math. No model call.
• 5 PM ET — Late-Edge Check. Sonnet does a focused pass — anything material change since lock-in? 95% of days: nothing new. Occasionally one ⚡ LATE ADD joins the board.
• 11:30 PM ET — Grader. ESPN finals come in. Each pick marked WIN / LOSS / PUSH. Ladder counter updates. Units P/L computed. Autopsy fires on every loss.

The Autopsy

Every loss gets a Claude post-mortem and is classified as one of:

• DATA — the data was wrong, missing, or stale. The pick might've been right with better data.
• MODEL — the reasoning was flawed. Wrong thesis, bad odds compare, ignored a factor.
• VARIANCE — the pick was sound; the outcome was unlucky.

Loss patterns suggesting new validator rules are queued for review. Rules are never auto-installed — humans decide what becomes law.

The Truth

The W-L record up top is what the boys see. The truth signal is the per-pick closing line value (CLV) — the gap between Scott Bot's price at publish and the market's closing line. Long-run positive CLV proves real edge. Win rates fluctuate over small samples. CLV doesn't lie.