How the engine works.

This page is the contract behind every slip we publish: what is modelled, how a slip qualifies, how it is staked, how it is settled, and what we will never claim. The proof for all of it is the public ledger.

The engine is deterministic. There is no machine-learning model to drift, no random seed to cherry-pick — the same inputs always produce the same outputs, and every read is reproducible after the fact.

Football

Goal markets come from per-side Poisson models seeded by each team's season scoring and concession rates. Volume markets — corners, shots on target — use negative-binomial distributions over those rates, which respect the over-dispersion real corner counts actually show. Player markets are per-90 rates behind a minimum-minutes floor: below it, the player simply has no read.

Above the per-market models sits a fixture-level fitness score built from four pillars — recent form weighted newest-heaviest, venue-aware points-per-game ability, how concentrated a team's outcomes are (predictability), and match context. A fixture has to clear that score before any of its markets are even considered.

Tennis

Tennis runs a hierarchical serve/return model. Each player's serve-points-won and return-points-won percentages on the surface combine against the tour average into a single point-win probability on serve. From there, point → game → set → match is a closed-form recursion — not simulation — and match winner, set scores, and total games all derive from the same hierarchy. Singles only, ATP and WTA, partitioned hard from football: no cross-sport slips, ever.

A slip qualifies by clearing probability floors at three levels: the fixture-fitness floor, a per-leg model-probability floor (set stricter for the premium 4-leg shape), and a joint-probability floor on the slip as a whole. Floors are hard gates, not soft penalties — a near-miss is a miss.

Odds never enter selection. Legs are ranked on model probability alone; prices attach afterwards, for display and settlement. A juicy price cannot argue a weak leg into a slip, and a short price cannot argue a strong one out.

The corollary: if nothing clears the floors, nothing is published. Quiet days are published as quiet — no padded slate, no junk legs to fill a schedule.

Every slip is staked at a flat 1.0 unit. This is enforced in code — the stake field in our system literally cannot hold another value. There is no martingale, no doubling after a loss, no “confidence” sizing. Slip shapes are capped per product (1-leg singles, 2-leg doubles, 4-leg multis, and the 8–10-leg cash-out rollup), and the ROI we report is net units returned per unit staked under that flat discipline.

One deliberate exception: football singles are generated continuously — often many per fixture — as an unstaked learning and calibration track, and they never enter the staked ledger. The football ledger records the 2-leg core shape and upwards, each at flat 1.0 unit. Tennis singles remain staked products and settle on the ledger like every other slip.

Prices are locked at the bookmaker's pre-kickoff line — Bet365 first, with a fallback chain — in the run-up to kickoff. What the ledger pays on is the price you could actually have taken, not a synthetic estimate; where a provisional price is later corrected to the real closing line, the original is kept for audit.

After full-time a settlement job grades every leg from the final match data. The rules are standard sportsbook treatment:

• —any leg lost → the slip settles lost, −1.00u
• —all legs won → the slip pays the combined odds of the surviving legs, minus the stake
• —a void leg drops from the slip and the remaining legs decide at recomputed odds; a fully void slip returns its stake, 0.00u

Every settled slip lands on the public ledger and stays there. The ledger is append-only: losses are shown as plainly as wins, never edited, never deleted, never filtered out of the headline numbers.

• —No guaranteed returns. Betting involves losing runs. The ledger shows ours, in full.
• —Signals are model reads, not certainties. A leg the model puts at 60% loses four times in ten — that is the read working, not failing.
• —Small samples are small. Per-product records start thin and we publish them anyway, with the sample size next to the percentage. We will not dress a 20-slip record up as a track record.
• —The model is not quietly re-fit. Calibration against settled results is monitored continuously, but drift raises an alert for a human — the model is never auto-tuned to flatter the record.
• —18+ only.We publish model reads; we don't place bets. Never stake what you cannot afford to lose — responsible gambling.