CLUTCHTIPTHE AI SPORTS ALMANAC
For sportsbooks, trading desks, media and funds

License the engine.

Calibrated player-prop pricing at scale, an independent second opinion on your own lines, and a public ledger that grades every number we publish. Market-blind by design, so its errors aren't your errors. Not tips — numbers you can price from.

1. Player props, priced at scale and calibrated in public

A line and a σ for every rotation player on every NBA slate — points, rebounds, assists, threes — priced from minutes, role and matchup with no reference to any posted line. Graded against the box score every morning in the open ledger:

StatActual inside published 1σIdeal
Assists76.0%68.3% = perfectly calibrated; above it means the stated σ is conservative
Points73.7%
Rebounds70.2%
Threes66.1%

4,911 graded lines, 22 Oct – 2 Nov 2026; 95–96% inside 2σ; realised error at or below the stated σ on every stat. Live figure in nba-calibration.json.

Thousands of lines a day is the product books find hardest to staff and the one growing fastest. This is a feed, not a pick sheet: delivered before tip-off as JSON, with the σ so your own risk tools can use it.

2. An independent second opinion on your lines

The engine never sees a price. That makes it the one input a desk can't build in-house: a probability formed without your line, so where it disagrees with you it carries information rather than echoing you. The divergence ledger joins your closing line to our number after the game and grades every disagreement — where we were shorter, where we were longer, who was right.

Already published for two full seasons: NRL 2026 (engine 69.0% v closing favourite 62.6%; 24 of 35 contrarian calls won at an average $2.34) and AFL 2026 (38 contrarian calls audited). For the NBA it runs against the engine's own odds archive and can be produced against your lines inside the trial.

3. On the roadmap: correlated same-game pricing

The player distributions and the game outcome come from the same inputs, so the correlation between legs — a player's points line and his side winning — is a derived quantity, not a new model. Same-game multi pricing with honest correlation is the piece most engines can't do and the piece books pay most for. Being built on the NBA props stack now; Same Race Multis on the racing side already run on the same joint-distribution machinery.

Why trust the numbers

Because every one of them is graded where you can see it. The win-probability record is the credibility under the products above, not the product:

LedgerRecordContext
NBA 2026–27, every game, pre-tip90 – 40 (69.2%)Favourite ≈67%; best published model 68.1%; ELITE 40 – 9 v best published band 75.7%
NBA 2025–26, every game425 – 152 (73.6%)Two seasons 515 – 192 (72.9%), 707 games
NRL 2026, every game141 / 204 (69.1%)Ahead of every expert panel winner; +9.0% flat v favourite −10.3%
AFL 2026, every game143.5 / 199 (72.1%)Grand Final called and won
Furlong, every Australian TAB race1 in 4Everest, Cox Plate, Melbourne Cup winners on top, no odds in the model

As at 8 Nov 2026. NBA and racing figures refresh nightly from the ledger; NRL and AFL are closed seasons audited against closing lines. Sources: results-nba, results-nrl, results-afl, furlong, record.json.

The calibration certificate

A strike rate tells you nothing about whether 65% means 65%. So the ledger publishes its own calibration, recomputed nightly from the same sheet the record page reads: Brier score, base-rate skill, log loss, expected calibration error, a reliability curve by confidence band and tier, and the props σ rate by stat.

StatedGamesWon
50–54%1361.5%
55–59%4763.8%
60–64%3767.6%
65–69%1478.6%
70–79%3281.2%

MAX AI's own confidence on every graded 2026–27 NBA game to 8 Nov (145 games, Brier 0.209). Every band has won at or above what it said: the stated probability is a floor. The published-tier view and the fitted-model era are reported separately in the file so nothing hides in an average.

Delivery, provenance, terms

Delivery: keyed JSON API or push to your endpoint; the NBA card lodged by 09:30 ET each game day with a pre-tip refresh that re-reads and stamps any game whose availability changes; a settlement file each morning; history back to the first lodged card; a written revision policy — nothing is deleted, a changed call carries its earlier read.

Provenance: licensed event data (BallDontLie, SportsDataIO, Rotowire, Punting Form); no scraped odds, no bookmaker affiliation, no bets taken, no odds shown to consumers. Method: proprietary and not published; every output and every grade is. Operator: an Australian proprietary company; ClutchTip is an analytics service, not a wagering service. Exclusivity by market or sport is available.

How a trial works

Thirty days, no fee, no obligation. We deliver the NBA card and props file to your desk every game day before tip-off — the same numbers our members see, plus the model prior and the σ — and you grade them against your own closing lines. At the end you hold a divergence ledger on your numbers and a calibration curve on ours. If it earns a place in your stack, we agree a licence. If it doesn't, you keep the ledger.

Request a trial How to cite the record

Media and broadcast: the nightly card — pick, tier and read — is available for editorial use with attribution, the public ledger as source of truth. Ask about a season licence.