SplitWinnerPublicly verifiable — you never have to trust us.
Institutional-grade models, a public immutable ledger, Bitcoin-anchored timestamps. Skips disclosed, not hidden. Built for sportsbooks — one git clone stands between any claim on this page and your own verification.
Live
§ 01The thesis
In this category, faked records are the norm — losers quietly deleted, winners backdated, every claim a sales pitch. We refused. Every live prediction is cryptographically committed the moment it's made, anchored where we can't reach it, and publicly verifiable by anyone — without trusting us.
01
Sports data pulled daily from canonical sources.
02
Per-sport models, trained with cross-validation, registered with held-out metrics.
03
Calibrated probabilities, conformal sets, computed once per day.
04
Each day's predictions committed to the public ledger before kickoff. Daily proof, upgraded to a Bitcoin attestation.
05
Through a signed, versioned API into the mobile app; and into the public GitHub mirror anyone can run verify.py against.
No marketing. Just the receipts.
§ 02The public alpha
From May 19 to June 25, 2026, the system ran its public alpha — every pick hashed into an immutable ledger and anchored before kickoff, skips disclosed, nothing deleted. The alpha existed to answer one question: does the backtest survive live games? The record is sealed, archived, and independently verifiable. Compare the columns yourself.
Sealed alpha · anchored · May 20 – Jun 25
Every figure on this card is computed only over picks in the sealed public ledger — recomputable by anyone from the repo and contract data. The sealed alpha record
Backtest · historical
Historical replay. Never anchored, never blended into the live record. Open here so you can see what the model has been doing.
The official SplitWinner ledger opens with its first anchor — MLB first, more sports as their seasons start. The alpha chain is closed; the two are never mixed.
Outside the sealed ledger — disclosed, never blended
The live pipeline has run since 2026-02-16. 926 settled picks sit outside the sealed ledger — most predate public anchoring, the rest fell on days without an anchor. They carry immutable database timestamps: verifiable to us, not verifiable to you — which is exactly why they are excluded from the sealed card above and from every anchored claim on this site.
Wins / losses, observed win rate, and units P&L per model — the FULL live record, including picks settled before public anchoring began. Source: /v1/quant/model-stats. 16 models with settled samples.
Live · refreshed hourly
| Sport | Model | Category | Sample | W–L | Win rate | ROI (u) |
|---|---|---|---|---|---|---|
| mlb | MLB ML 1 | moneyline | 438 | 265–173 | 60.5% | +35.0 |
| mlb | MLB ML 2 | moneyline | 276 | 172–104 | 62.3% | +57.0 |
| mlb | MLB SP 1 | spread | 381 | 230–151 | 60.4% | +26.9 |
| mlb | MLB SP 2 | spread | 493 | 296–197 | 60.0% | -26.9 |
| nba | NBA ML 1 | moneyline | 232 | 180–52 | 77.6% | +221.9 |
| nba | NBA ML 2 | moneyline | 228 | 176–52 | 77.2% | +205.4 |
| ncaab | NCAAB ML 1 | moneyline | 1075 | 779–296 | 72.5% | +107.4 |
| ncaab | NCAAB ML 2 | moneyline | 1101 | 803–298 | 72.9% | +117.3 |
| ncaaf | NCAAF ML 1 | moneyline | 167 | 127–40 | 76.0% | +45.2 |
| ncaaf | NCAAF ML 2 | moneyline | 201 | 149–52 | 74.1% | +41.5 |
| nfl | NFL ML 1 | moneyline | 70 | 54–16 | 77.1% | +47.3 |
| nfl | NFL ML 2 | moneyline | 53 | 42–11 | 79.3% | +51.3 |
| nhl | NHL ML 1 | moneyline | 103 | 66–37 | 64.1% | +94.6 |
| nhl | NHL ML 2 | moneyline | 79 | 53–26 | 67.1% | +74.3 |
| nhl | NHL SP 1 | spread | 616 | 427–189 | 69.3% | -71.6 |
| nhl | NHL SP 2 | spread | 650 | 448–202 | 68.9% | +74.6 |
Coverage
Every pick lands in the same public ledger, across every sport we cover. Live and backtest counts kept separate — same model, two records, never blended.
Sports covered
6
Live predictions
4,738
Live conformal
996
Backtest predictions
17,043
Backtest conformal
5,167
Datasets / sport
2
§ 04Verify yourself
Three steps. Pure-stdlib Python, no toolchain. Clone the public ledger, run the verifier, check the Bitcoin attestations against the dates we claim. Every verifier release is registered and hash-checked in CI — and the sealed alpha record verifies the same way, forever.
git clone https://github.com/SplitWinner/audit_trailcd audit_trail && python3 verify.py vectors && python3 verify.py chainVendor questionnaires answered before they're asked.If we vanished tomorrow, the proof would still stand.
§ 05Track record
Three records, never blended: the anchored slice the sealed ledger can back, the full live record since February 2026, and the historical backtest. Same model, same code, same seeds. If the numbers are honest, the columns should rhyme — judge that yourself.
How these are computedLive · /v1/quant.live + /v1/quant.backdate · refreshed hourly
Showing backtest record
Win rate
57.1%
n = 70
High-confidence wins
83.3%
≥70% confidence
Brier score
0.2424
lower is better
Calibration error
0.0963
lower is better
| Metric | Aggregate | mlb | nfl | ncaaf | nba | ncaab | nhl |
|---|---|---|---|---|---|---|---|
Win rate observed win rate on settled picks | 57.1% | 55.4% | Sep '26 | Aug '26 | 83.3% | — | 50.0% |
High-confidence wins win rate when model confidence ≥ 70% | 83.3% | 50.0% | Sep '26 | Aug '26 | 100.0% | — | 100.0% |
Brier score mean squared probability error | 0.2424 | 0.2498 | Sep '26 | Aug '26 | 0.1570 | — | 0.2548 |
Calibration error expected calibration error | 0.0963 | 0.0868 | Sep '26 | Aug '26 | 0.1564 | — | 0.2212 |
Avg confidence mean predicted probability on scored picks | 64.6% | 64.0% | Sep '26 | Aug '26 | 69.6% | — | 65.0% |
Sample size settled predictions counted | 70 | 56 | Sep '26 | Aug '26 | 6 | — | 8 |
What each number should be compared to. Live-mode only — backtest never blended in.
break-even at −110 vig
The bar a sports-bettor has to clear to break even paying standard juice. The model publishes only when conformal coverage holds — abstention pulls this above the break-even line by design.
random-guess Brier on a binary
Mean squared error between predicted probability and outcome. Proper scoring rule — penalises overconfidence. 0.25 is the random-guess baseline; sub-0.21 is sharp.
typical ML models
Expected calibration error: the gap between what the model says and what it hits. Sub-0.05 is well-calibrated; sub-0.02 is rare. When the model says 70%, it hits close to 70%.
no edge, no return
Per-bet Sharpe over the live equity curve. Flat-bet (no Kelly) so leverage doesn't inflate it; the signal is the signal. Above zero = positive expectation.
Predicted probability vs. observed win rate. The diagonal is perfect calibration.
n = 70
Flat-stake units, settled picks only.
Max drawdown: -12.3u · Sharpe per bet: 0.096
ROI grouped by the model's predicted edge over market.
n = 70
§ 05For sportsbooks
The engine prices every game the way a sharp does — and shows its work. Put it beside your trading desk: see where your line disagrees with a calibrated, publicly audited signal before the money does. Delivered as an API feed your quants can interrogate, and a ready terminal your desk can run.
2026 beta cohort
We're onboarding a limited beta for the 2026 seasons — MLB live now, football and the rest as they open. Full feed, sandbox to production in days, a fixed seasonal price quoted in writing, direct access to the operator. When the season ends you hold your own settled record of what the signal was worth — the same ledger receipts we publish.
§ 06Products
The Feed is the integrator surface — the engine on the wire, sandbox open today. The Terminal is the same signal as a ready-to-run app, white-labeled for your desk. The Analyst is the natural-language layer on the full audited record. Same engine, same hashes, same anchors.
SplitWinner
Today's slate priced and graded, per-pick intel, and the audit receipt behind every call — in an app your traders open at 8am. White-labeled for your book, distributed privately to your team, never the public store.
ExploreSplitWinner GPT
A purpose-built LLM grounded in the full audited record. Ask why a model called a game, how a sport's been trending, what the record actually shows — every numerical answer cites the exact audit_trail anchor it was computed from.
ExploreSplitWinner API
The engine on the wire. The sandbox is free and self-serve — every endpoint mocked plausibly-live, integrate against the full surface in under an hour, no NDA. Production is direct contract: pull-and-cache, you serve your stack from your own infrastructure, ACH or wire. Built for trading desks, risk teams, integrity vendors, and leagues.
Explore§ 07Questions
Plain answers to the questions a skeptic asks first.
No. No honest prediction service can. Sports outcomes carry real variance — anyone who guarantees is not being honest with you.
This is the right question to ask — the industry is full of deleted losses and backdated winners. We built the system so we can't quietly delete a loss or backdate a win. Database triggers reject every UPDATE and DELETE on the predictions ledger. The day's predictions are hashed and committed to a public GitHub repo before the games start, then anchored to Bitcoin.
The full mechanism + daily reproducibility CIA pick is a person's opinion — easy to selectively remember, easy to forget. A prediction is a model-generated probability with a confidence score, logged and scored against the outcome. No cherry-picking, no story — just results measured over a real sample.
Yes. So we don't show them as live calls — we show them next to the live record, labeled, with the model's behavior in both. A backtest run on past games is honest characterisation; calling it a live call is fraud. We never blend the two.
Because we're new. Daily anchoring launched May 2026 — every live pick since is in the public ledger. We could pad it with our backtest history, but that's the trick we built this to prevent. The short live record is the honest signal.
See the May 2026 boundary on the verification timelineWe have a Kalshi entity account that a team member trades using the predictions plus their own judgment — not a mechanical bet on every pick. That's our internal modeling test, not a public-proof channel. Beyond that: real edges in sports markets are small after vig and reflexive, sportsbooks throttle sharp accounts in weeks, and posting every prediction to a public order book would telegraph the pick to anyone watching. Selling access scales without those constraints. The proof channel is the audit-trail anchor.
Pick the lane that fits — a short note when the drop or methodology changes, or a direct line to the enterprise track.
Build against the API
A limited cohort of desks, a fixed seasonal price in writing, the full feed. Start in the free sandbox — no card, no NDA, integrated in under an hour — and when it holds up, the beta is a direct conversation with the operator.
Operator briefing
Methodology changes, incident disclosures, cohort openings — the operator's notes, a few times a season. Company and role optional, but they get you routed like a desk, not a list.
No marketing. Just the receipts.