The first-half MLB edge map.

The All-Star break is the natural midpoint of the season and the right moment to audit what the model has done with the first 90 or so games per team. Half a season is enough sample to see the shape of the signal. It is not enough sample to declare victory or defeat. Both of those framings would be wrong.

What the first half produced.

The MLB model fired signals on roughly one in four available games across the first half. That rate is consistent with the discipline filter operating as designed. About 75 percent of MLB games do not carry an edge that clears the threshold. That is the math of a sharp market.

Closing line value across published signals was positive. The exact figure lives on the public ledger. The shape of the CLV distribution is more useful than the headline number. Most signals beat the closing line by a small amount. A smaller number beat it by a larger amount. A few missed. The distribution is what a real edge looks like.

Where the first half worked.

Bullpen game totals.

The piece from earlier this summer described why the market treats bullpen games as a single category and the model treats them as nine pitching decisions in a row. The first half validated this. Bullpen game totals were the highest-CLV category in the entire MLB book.

Wind-affected park overs.

The weather piece described the narrow band of conditions where wind actually moves the line. The model fired roughly two dozen signals in that band across the first half. They produced the second-highest CLV of any first-half category.

Specific starter-versus-lineup mismatches.

When a starter's handedness, breaking-ball usage, and recent form combined poorly against a specific lineup's tendencies, the model found edge the market underpriced. This is a multi-variable read, not a single-input one, and it is exactly the kind of edge the market handles individually but not in combination.

Where the first half struggled.

April lines on coastal pitchers.

The model overweighted spring training and brief early-season data on West Coast and Northeast starters whose mechanics were measurably different in cold weather. The closing line audit caught these. CLV in April was meaningfully lower than CLV in May or June.

Run line bets generally.

The CLV thesis on MLB run lines has been weak for two seasons running. The structural issue is that run line movement is mostly juice movement, not spread movement, and our CLV calculation does not capture juice as cleanly as it captures spread. This is a known limitation. The first half did not produce a fix. We are still working on it.

What the second half changes.

Two concrete adjustments.

The cold-weather starter input now uses a temperature-adjusted recent form window instead of a fixed lookback. Pitchers who throw in cold weather during April get evaluated against their cold-weather history, not their full-career baseline. This is a small change with a meaningful expected impact on April and early-May signals next season.

The run line filter tightens. The model will publish fewer run line signals in the second half. The ones that do publish will carry a higher edge threshold. This is the conservative response to a measurement problem we have not yet solved.

What does not change.

The architecture, the discipline filter framework, the pass-day framing, the closing line audit. The brand thesis is unchanged. CLV is the institutional scoreboard. A first half that produced positive CLV is a first half where the system worked, regardless of how individual nights felt.