THE SWEAT LOCKER

Most sports analytics apps run your question through an AI and call it a model. We built an actual pipeline.

Every day, across every sport we cover — MLB, NFL, NCAAF, UFC live now, with NBA, NHL, and NCAAB rolling in — our system pulls real data. Pitcher Statcast metrics. Snap counts and EPA per play. Offensive line vs defensive line matchups. Fighter striking and grappling profiles. Net efficiency and pace. Park factors, weather, umpire tendencies. Injury reports and rest days. And on top of it all, live market data: opening lines, per-book line movement, sharp money splits, and public bet percentages.

That data runs through the pipeline throughout the day — before locks and after grades — writes to a live database, and feeds every score, badge, and pick you see in the app.

Then four models argue.

Every game and every prop gets checked by an ensemble panel, a logistic regression shadow, a 10K-simulation Monte Carlo, and a cohort match against historical situational patterns. When they agree, it ships as PRIME. When they disagree, we downgrade it — or fade it. You see the disagreement on the card. No black box.

Then Jerry reads the output.

Not the other way around. Jerry doesn't guess picks and dress them up with stats. Jerry reads what the four models actually produced and translates it into plain language — the signal, the counter-signal, the confidence, the risk. If the model says PASS, Jerry says PASS.

Then we grade everything.

Every pick published, every day, logged with the model output that produced it. Wins in green, losses in red, on the home screen the morning after. Every cohort audited daily — if a tier stops hitting, we see it within days, not months. When a play type breaks (like our OVER-lean signal that went 0-10 in April), we pull it, publicly. When a tier overcorrects, we tell you. No manual scorekeeping. No retroactive edits. No cherry-picked highlight reels.

The model gets a report card every morning. Users see it before we do.

We don't take affiliate money from any sportsbook. Every dollar of revenue is from your subscription. Independence is the product.

The deepest baseball coverage in any consumer app. Every game runs through four models — ensemble panel, LR shadow, Monte Carlo dissent, cohort match — with a Playbook scoring layer on top. Every pick shows where the models agree and where they disagree.

• xERA + first-inning splits — not just season ERA. How does this pitcher perform in the first inning specifically, against this lineup's handedness profile?
• Platoon advantage — batter handedness vs pitcher handedness, weighted by lineup position
• Pitcher contact profile — BAA, xBA, hard-hit % allowed for prop analysis
• Bullpen quality — ERA differential as a late-game signal
• NRFI model — proprietary first-inning scoring, published only on extreme edges with PRIME conviction
• Park factors, weather, umpire tendencies — every environmental variable that moves lines
• Spread projection — our own run-differential model vs the posted line
• Prop Jerry — every prop card shows L5/L10 form vs the line, Signal Coverage pill, and Playbook verdict (BACK/FADE/PASS)

The biggest betting market in America. Our NFL model shipped Week 1 of the 2026 season, tuned against a 66.7% sides backtest window. Every game processes:

• EPA per play — offensive and defensive efficiency baselines
• Target share + air yards — receiver prop signal
• Situational edges — bye week, dome team in cold weather, timezone fatigue, short week
• Red zone efficiency — TD prop grading
• Weather modeling for outdoor stadiums
• Line + rotation talent signal — Madden 27 Top 100 as roster-talent prior, OL/DL average weight as trench-play input
• Same 4-model architecture as MLB — ensemble, LR shadow, Monte Carlo, cohort match — with the same PRIME/STRONG/LEAN conviction tiers

Week 2 live. Hit rates + full audit visible in-app.

College football, modeled the same way as the pros. Phase 2 shipped August 2026 with an SP+ inspired efficiency layer, returning production adjustments, and 7 published fade rules.

• SP+ inspired efficiency — offensive and defensive baseline per team
• Returning production — roster continuity as a Week 1-6 signal
• Roster physicality — offensive/defensive line average weight, class-year distribution
• Situational spots — trap games, lookahead spots, revenge angles
• Weather for outdoor games
• 7 fade rules — published patterns that consistently produce losing sides (public trap zones)
• Same 4-model architecture — plus a Playbook layer that scores each pick by signal source

Week 3 tomorrow. Multi-week backtest visible in-app.

Cage-side model with real fighter profiles, not scraped tale-of-the-tape charts.

• Comprehensive fighter database — strike, takedown, control, and finishing splits with full fight-by-fight history
• Proprietary 4-model XGBoost ensemble — predicts winner, method (KO/SUB/DEC), goes-distance, and round
• Isotonic calibrator — probabilities calibrated against 2-year fight history, not raw model output
• Opponent matchup engine — Jerry sees both fighters and their history together, not one in isolation
• Web search integration — cross-references analyst consensus against model output
• PRIME / STRONG / LEAN conviction tiers per fight, plus per-method side bets

Regular season starts October 21. We ship LEAN-only picks through the calibration window while the model builds a real 2026-27 sample. STRONG and PRIME tiers activate once we clear ~20 games of sample per team.

• Net rating, defensive rating, eFG%, pace, OREB%, TOV% — the six pillars
• Elo model shipped August 2026 as a supplementary signal
• Recency-weighted player averages for prop analysis
• Opponent defensive adjustment by category for prop grading
• Playoff mode — series context, elimination urgency, home court earned
• Injury report integration with real-time status updates

Live audit visible in-app during calibration.

Regular season October 2026. Our NHL stack runs on the free NHL API + MoneyPuck data — no paid tier required for consumers, no vendor risk for us. Model activates for regular season.

• Team pillars — expected goals for / against, high-danger chance rate, PP / PK efficiency
• Elo model shipped August 2026 as a supplementary signal
• Goalie factor — starter save % vs opponent shot quality
• Rest & travel — back-to-backs, road trips, timezone shifts
• Scoreboard integration with 1,335-game historical backfill
• Grader + resolver ship in late September for preseason grading validation

Foundation deployed May 2026. Launches with a 4-lens efficiency architecture on Day 1, with V4 model targeted for December 3.

• KenPom-inspired efficiency model — 4-factor architecture (eFG%, TO%, OR%, FT rate)
• Tempo + strength of schedule normalization
• All 365 D1 teams normalized through our alias table
• Roster continuity — returning minutes as an early-season signal
• Player class-year + height as physicality inputs
• PRIME / STRONG / LEAN conviction tiers matching every other sport's system

Daily picks activate when regular season tips off.

Jerry — Your Pipeline Analyst

Most "AI-powered" sports apps run your question through a chatbot and call it a model. Jerry doesn't search the internet and guess.

Jerry reads the pipeline output — the ensemble prediction, the LR shadow, the Monte Carlo dissent, the cohort match, the sharp-money split, the model consensus — and translates it into plain language. Signal, counter-signal, confidence, risk. Three sentences. No fluff. No tab-switching.

If the model says PASS, Jerry says PASS. If four models agree, Jerry tells you why. If they disagree, Jerry names the disagreement.

Jerry lives on every pick in the app, across five surfaces:

🎯 POTD Read — The Play of the Day's narrative. One pick per day, cross-sport. Jerry names the edge, the counter-signal, and the conviction. Locked before games start.

📇 Sweat Card Read — Your daily dashboard's per-pick summary. Jerry writes the one-liner for every card on the deck.

💎 Prop Jerry — On every prop card: L5/L10 bar chart vs the line, Signal Coverage pill showing which data sources fired, Playbook verdict (BACK / FADE / PASS), and a plain-language read on why. PRIME / STRONG / LEAN tier with a 0-100 conviction score. Currently ships across MLB, NFL, UFC live; NBA calibrating; NHL and NCAAB preseason.

🎲 Daily Degen — A 3-5 leg parlay composed from the slate's highest-conviction signals. Same pick for every user. No personalization, no upsell. Either the slate had it or it didn't.

🐕 Dog of the Day — One underdog moneyline per slate where the model sharply disagrees with the market. Triggered only when the model-vs-market gap clears our threshold — most days there isn't one, and we don't force it.

Then everything gets graded.

Every Jerry-narrated pick is logged with the model output that produced it and auto-graded the morning after — via MLB Stats API, official NFL / NCAAF box scores, UFC fight results, BDL for NBA, and the NHL API. Wins in green, losses in red, on the Receipts calendar. No manual scorekeeping. No retroactive edits.

Jerry is powered by Anthropic's Claude AI model. Prompts are not stored or used for training. Full disclosure in our Privacy Policy.

Every game gets a Sweat Score from 0 to 100. Scores at 68 or above are Prime Sweats — plays where multiple independent signals stack in the same direction.

The score isn't vibes. It's a weighted blend of five signal buckets:

· Model Mismatch — the 4-model panel vs the posted line · Situational Edge — rest, travel, weather, spot, splits · Line Trajectory — where the number opened, where it is, who moved it · Sharp Signal — sharp money %, per-book divergence, RLM · Market Efficiency — how tight or loose the market is on this specific game

The weights shift by sport, based on how much we actually trust our pipeline there:

SportWhat dominates the scoreMLBModel leads — deepest pipeline, 4-model architecture, situational + line-trajectory reinforceNFLBalanced — model + market + sharp money weighted evenly, situational edge (weather, dome, rest) heavier for outdoor gamesNCAAFSP+ inspired model leads, roster continuity + physicality + situational fade rules reinforce, sharp money as tiebreakerUFCSharp signal leads — MMA lines move fastest on sharp money — model + market as checkNBAModel + situational lead once regular season sample matures; calibration-weighted through OctoberNHLElo model + market signal balanced, sharp money as tiebreaker, situational for back-to-backsNCAAB4-lens efficiency model + strength-of-schedule lead once regular season tips off (Nov 3)

The score is compressed, not inflated.

Most apps tune their headline number up so every game feels like a Lock. We do the opposite. Our normalization curve maps a raw 75 down to a displayed 68 so the Prime Sweat threshold genuinely means rare and stacked. Fewer Prime Sweats. Higher hit rate when they fire.

Your daily dashboard. One place. The whole slate at a glance.

The Sweat Card composes throughout the day as the pipeline runs — final lineups, umpire assignments, weather, and closing lines feed in, and the card refits automatically. When you open the app, this is the first surface you see.

What's on it:

· Yesterday's Record — W-L for POTD, Sharp Card, and every graded pick. Green wins, red losses. Right on top. · Today's Play of the Day — the highest cross-sport conviction pick, with Jerry's read. · Dog of the Day — when the model + market diverge enough to publish one. Some days it's blank. · Top Curated Plays — the day's PRIME and STRONG picks across every live sport. · Daily Degen — a 3-5 leg parlay built from the day's highest-conviction signals. · Upcoming Events — tomorrow's UFC card, tomorrow's MLB slate, next weekend's football. · Live audit hit rates — the tier record backing every card, updated daily.

No hunting through tabs. No decision fatigue. Open the app, see the day.

We track everything publicly because we believe in the model.

Every pick is logged with the model output that produced it. Every cohort audited daily across rolling 7-day, 30-day, and season-to-date windows. When a tier stops hitting, we don't quietly bury it — we downgrade or pull it, publicly.

• We pulled our OVER-lean badge in April after it went 0-10.
• We downgraded NBA Net Rating STRONG-tier to LEAN in May after it audited at 36% on 11 games.
• We paused teased_totals combos in August after 7 straight days of -3u.
• We killed chalk_prop_parlay in September after 30 days audited at -3.98u.

The misses are loud on purpose. Nobody else in this category does this.

What's tracked, all visible inside the Receipts tab:

🎯 Play of the Day — every POTD graded on the calendar. Green wins, red losses, gray pushes. Tap any day to see the actual pick. 120-day history.

🎴 Sharp Card — the daily curated deck's PRIME + STRONG record. W-L, unit P/L, rolling 30d + MTD + previous month. Every card that shipped, every day.

🧾 The Ledger — chalk parlay + teased combos record. Per-kind breakdown so you can see which combo types are hitting and which we've killed.

💎 Prop Jerry — conviction-tiered (PRIME / STRONG / LEAN) results by tier and by prop type. Hits, HA, KS, Outs, ER, BB — each broken out.

🐕 Dog of the Day — every underdog ML the model fired. Rolling W-L and pending count. Blank days included.

🐟 Ladder — one play per day, roll winnings 30 days. Full 30-day arc visible.

📊 Live tier calibration — rolling hit-rate audit per cohort, refreshed continuously as the pipeline runs. If a tier drifts, we see it within days, not months.

Auto-resolved via MLB Stats API, official NFL / NCAAF box scores, UFC fight results, BDL for NBA, NHL API. No manual scorekeeping. No retroactive grading. No cherry-picked highlight reels.

Past performance does not guarantee future results. This is analytics, not a guarantee.

The edge in sports betting isn't picking winners. It's finding spots where the market has mispriced something the model sees clearly.
A pitcher with a 5.2 xERA and a 5.8 ERA bullpen behind him facing a wRC+ 115 lineup — that's not a hunch. That's a pipeline flag that triggers a conviction grade before we ever check the odds.
The same shape plays out in every sport we cover:
· NFL — an offense hitting 0.15 EPA/play into a defense giving up 0.24 in the second half, playing outdoors in 22° with a 15 mph crosswind. The market prices the total off season averages. The model prices it off the weather.
· NCAAF — a road team with a 320-lb OL trench advantage over a defense that hasn't seen a Power-4 offense this year. The line moves off the ranking; the model moves off the trenches.
· NBA — net rating gaps the market hasn't fully priced when a team's L10 pace + shot profile shifts. The line lags. The model catches it.
· UFC — striker-vs-grappler matchups where the line prices the name recognition, not the stylistic mismatch. Sharp money moves first. The model moves with it.
· NHL — a back-to-back road team with a backup goalie against a rested opponent with elite high-danger chance suppression. The market treats it as a coin flip. The model doesn't.
· NCAAB (Nov 3 launch) — pace and 4-factor efficiency mismatches on games where the ranking tells one story and the underlying metrics tell another.
Then we grade it.
Every flag, every conviction tier, every miss — logged and audited daily, in plain sight. When a tier stops hitting, we pull it. When the model overcorrects, we say so. When four models disagree and we take the underdog anyway, you see the disagreement on the card.
That's the difference between The Sweat Locker and everything else.
More Data. Less Sweat.