The Lens is a model-driven MLB prop analysis tool. Every day it publishes its top pitcher strikeout looks and top home run looks, saves them with a predicted probability before games start, and grades them automatically against official MLB box scores overnight. Nothing is edited, deleted, or cherry-picked after the fact — the Model Record above is the complete history.

Three enforcement rules back that up: every pick carries a database creation timestamp, picks are never saved for a game that has already started, and postponed or suspended games are marked VOID — excluded from hit rates in both directions rather than quietly dropped or counted as misses.

For each starter, the model estimates the probability of reaching 4+ through 9+ strikeouts using a binomial distribution over his expected batters faced. The inputs, all visible on every pitcher card's K PROB tab:

• Season K/9, blended 70/30 with his last five starts so current form counts
• Opponent lineup strikeout rate vs the ~22% league average (whiff-prone lineups raise the number, contact teams lower it, capped at ±20%)
• Expected batters faced, from his median innings over the last five starts × 3.8 batters per inning

The graded threshold is 5+ strikeouts with 4+ innings pitched — a hit or a miss, no partial credit.

Each hitter pick stores the predicted chance of 1+ home run, built from: his HR rate regressed toward league average (small samples can't fake it), blended 60/40 with what his Statcast barrel rate says the rate should be (skill over luck), then adjusted for ballpark, the opposing pitcher's HR tendency, the platoon matchup, lineup slot (more plate appearances = more chances), and game-time weather — temperature and wind resolved against the park's orientation. A hit is 1+ HR.

Honest selection: a hitter only becomes an HR pick when his HR probability clears the model's floor. Hitters below it aren't dressed up as HR looks — the model offers their strongest honest prop instead: total bases (2+ TB), computed from regressed single/extra-base-hit rates walked over expected plate appearances. On a thin slate, fewer picks are published rather than weaker ones.

A model is calibrated when its predictions match reality: picks tagged 65% should hit about 65% of the time. The panels above compare what the model predicted against what actually happened, in buckets, with minimum sample sizes so a lucky week can't paint a pretty chart. The Brier score summarizes it in one number (lower is better). When the data shows a gap, the model's constants get tuned — publicly, on this page.

Late scratches, unannounced injuries, a manager's quick hook, weather that shifts after picks lock. And probability is not certainty: a 70% pick should miss three times out of ten — that's the model working, not failing. Nothing here is betting advice; it's analysis with a public scorecard.