SPORTS · LIVEGolfSoccerMLBUFCNFLWNBALABNBA · SOONNHL · SOONF1 · SOONNASCAR · SOON
THE STANDARD

Three rules.
No exceptions.

01
We publish probabilities, not predictions
You'll never see “Mexico will win” here. You'll see how likely we think each result is — and then you can watch whether the things we call 46% actually happen about 46% of the time. That's the only honest test of a number like ours, and we run it in public.
02
Timestamped before lock. Never edited after.
Every projection is written to a journal before the match kicks off or the field tees off. Grading runs against that journal — not against anything that could be touched up after the result. A track record you can't rewrite is the entire product.
03
Misses shown with the same prominence as hits
When we're wrong, it goes on the board next to when we're right. No cherry-picked screenshots, no quiet deletions. If a track record ever looks too clean, you should be suspicious — ours won't.

How grading works

Before every match or tournament locks, the day's projections are written to a journal file with a timestamp. Once official results post — typically within a few hours of full time — an automatic grader scores every journaled number against what actually happened and publishes the running record on Insights. The full journal — every entry, every timestamp, raw files included — is public at The Journal. The grader can only read the journal. It cannot see — and therefore cannot flatter — anything written after kickoff.

A real journal entry, graded

Journaled June 10, the day before the World Cup opener:

MEXICO v SOUTH AFRICA · JUNE 11 · ESTADIO AZTECA
Mexico win 75% · Draw 17% · South Africa win 8%
GRADED · FT 2–0 MEXICO · the 75% call landed

Those numbers were locked before kickoff and graded within hours of full time — the first line of the tournament's permanent record. Every match since has gotten the same treatment, hit or miss.

What counts as a pick

The model produces a probability for every match in every market. A probability is not a pick. A pick is a public commitment to a position, and we make it only when the model has found a material edge against the market— when our number and the market's disagree by enough to matter. “The model likes the favorite” is not an edge. The market usually likes the favorite too. The edge is the disagreement, and it has to be big enough to clear a bar we hold ourselves to privately.

That bar is deliberately high, and it's higher in some markets than others — tighter, more efficient markets demand a wider edge before we'll commit, and we hold the model to where it has actually earned trust. The exact triggers are ours; what we'll tell you is the principle: most of the board doesn't clear it. When the model only agrees with the market, or the match is a coin flip, there is no pick— and that restraint is the product, not a gap in it. The goal is a record that's small and sharp, not large and average. A quiet day with no pick is a signal too.

Pick types are kept separate, never rolled into one number — a model can read sides well and totals poorly at once, and averaging them hides that. And the picks record is tracked apart from calibration. Calibration measures the whole model: across every match, do the things called 60% actually happen about 60% of the time. Picks measure the result of the commitments made when the model saw an edge. Both numbers are public. Both are honest. Misses carry the same weight and the same red as hits, and at small samples the number is shown but no conclusion is drawn from it.

SEE THE BOARDREAD THE JOURNAL →