How we grade a pick — and judge ourselves.

The closing line is the truth serum.

The closing line is the price a market settles on right before kickoff. By then it has absorbed every injury, lineup, weather report and sharp bet — it is the single most accurate estimate the market ever produces.

So we don't measure ourselves against wins. We measure whether the price we took beats the price the market closed at. Consistently beating the close — positive Closing Line Value (CLV) — is the signature of a real edge, and it shows up long before profit does.

We strip the bookmaker's margin first.

Every quoted odds contains the bookmaker's margin (the vig or overround) — the built-in edge that makes probabilities sum to more than 100%. Comparing an AI estimate to a raw price would be comparing against a number that's deliberately shaded.

So before anything else we remove the margin to recover the market's fair, no-vig probability — using proportional de-vigging for three-way markets and the Shin method for two-way ones. Only then is a comparison honest.

A calibrated model grades every market.

An AI model analyses each match pre-game and outputs probabilities for every market it covers. The rules are strict, not vibes:

• Home, draw and away probabilities must sum to 100% — checked before anything is saved.
• Confidence is anchored to explicit thresholds (estimated probability, edge size, number of independent arguments) — not a gut feeling.
• Below a minimum edge, a pick is treated as noise and not reported.

Edge is model probability minus the fair market price.

Value isn't “this looks good”. It's a number:

A pick only qualifies when the model's probability is meaningfully higher than the market's honest price. Everything below that threshold is discarded before it ever reaches you.

Timestamped when taken, judged at the close.

• Each pick is logged at the best publicly available odds at the time, with a timestamp — no rewriting history after the result.
• After kickoff we compare that price to the closing line. Beating it, over many picks, is the edge.
• Duplicate and self-contradictory picks are excluded from the record automatically, so the stats can't be inflated by counting the same bet twice or by hedging both sides.

Calibration, intervals, and nothing hidden.

A model that says “60%” should win about 60% of the time. We publish the calibration curve (predicted vs actual), a Brier score, hit rate with Wilson confidence intervals, and running CLV — updated whether we're up or down.

We also run a formal gate: the edge is only considered proven when the lower bound of the CLV confidence interval clears zero over a large sample. Until then, it's a candidate, not a claim.