How the picks work.

No black box. Here's exactly what every number on a pick card means and what the AI actually researched before it shipped.

Expected Value (EV)

EV is the only number that matters long-term. It's how much you'd profit per $100 bet, on average, if you took that exact pick infinite times.

A +12% EV pick means the math says you make $12 per $100 staked over the long run. It does not mean it'll hit. Individual bets are coin flips with weighted coins. EV tells you the weight.

Why it can be positive at all: sportsbooks price picks based on public action and their margin (the "vig"). When public money piles onto a popular favorite, the books shade the line away from the true probability to balance their book. That gap — between true probability and book-implied probability — is the edge. The AI scans every leg in every game and only ships picks where the gap is in your favor by at least 2pp after calibration. Below that, no slate.

Confidence

The model's estimated chance the whole parlay hits. 65% confidence on a 2-leg parlay means the model thinks ~65 of every 100 of these would win.

Confidence drops fast as you add legs because probabilities multiply. Two 80% legs = 64% combined. Three 80% legs = 51%. Four = 41%. That's why most slate picks are 2-leg — the math punishes long chains, and we'd rather ship picks that actually feel achievable.

Conviction Tiers

Every pick on the slate has positive EV by definition (no negative edge gets through). But not every positive-edge pick is created equal. We surface a conviction tier so you can see at-a-glance how many signals agree:

🔒 LOCK

All six signals aligned. 5+ pp model edge, LLM verifier actively endorsed (not soft-pass), 2 legs or fewer, ≥55% combined confidence, no hedge words in the AI review. Highest conviction tier the slate produces.

STRONG

3+ pp model edge with the LLM keeping or staying neutral, on a parlay of 3 legs or fewer. Solid pick, just doesn't hit every Lock criterion.

SOFT

Math says positive EV but the LLM verifier flagged caution — usually a structural concern (one weak leg, sketchy stack, marginal edge). Still on the slate because the EV gate cleared.

No tier is a guarantee. -300 favorites lose ~25% of the time. A "Lock" on a slate could lose tomorrow. The tier tells you how many signals stacked in your favor before the ball was thrown — that's it. Anyone selling you a guaranteed lock is selling you something.

What the AI actually researches

Per leg, before any pick gets scored:

Model probability vs book. Logistic regression trained on historical results estimates the true win rate. Compared against book-implied probability across 4+ sportsbooks to find the best price.

Recent form. Last 5 games, win streaks, head-to-head history when relevant.

Injury & lineup. Active injury reports for both teams. Backup goalies (NHL), key player out (NBA), bullpen workload (MLB) all flagged before pricing.

Pitching match-ups (MLB). Starter ERA, xERA luck gap, opposing lineup xwOBA, barrel rate edges. Surfaces pitchers who've been unlucky vs ones who've been lucky.

Sharp money signals. Public bet % vs public money % — when 70% of bets are on Team A but only 50% of money is on Team A, sharps are taking the other side. We track that lean.

Line movement. Where the line opened vs where it closed. Movement against the public usually means sharp action.

LLM sharp-eye review. After all the math, every candidate parlay goes through a Gemini pass that judges it like a sharp human would: is this structurally sound, are any legs sketchy, does the claimed EV pass a sanity check? The verdict (keep/soft/skip) and one-line reason show up on every card.

What gets a pick cut

EV under +2% after calibration — no edge, no slate.

LLM verifier returns "skip" — structural problem.

Diversity filter — too similar to a pick already on the slate.

Sport not currently priced — NBA + UFC are excluded from parlays until the model is retrained on them. UFC has its own dedicated picker at /ufc.

The honest part

Sports betting is a -EV game by default — books take a cut on every line. The only way to win long-term is to find the lines they mispriced and bet those, then size your bets so variance doesn't kill your bankroll. That's what this is.

Even the best signal stack will lose individual bets. Variance is real. The slate ships picks where the math is in your favor — your job is to bet sized stakes consistently and let the math play out over hundreds of bets, not three.

See today's slate →