good-sport · the ledger
Claims we make, graded as evidence lands.
Learnings from building a sports model that grades its own picks. Entries that make a claim get a stamp — and we come back and grade it, the same way the model is graded. No highlight reel; the losses stay up.
live gates
Every family the model bets has a pre-committed keep/kill rule. These are the fuses, filling as bets grade — when one fires, it becomes an entry below. Numbers refresh with each night's grading.
- tennis re-entry 31 / 100 · +13.6pp tripwire FIRED 8/15 at n=35 (−14.1pp) — benched; post-fix bets earn re-entry at n=100 if ≥ +2pp. ~6 bets/wk, so this is a December question — and the serve-stat coverage gap has to close before the number means anything
- batter hits bar 1263 / 400 · +4.4pp PASSED 8/1, 8/15 and a third time 9/1 (+4.1pp post-recal window, n=723) — full weight; same bar re-checked each audit
- batter strikeouts bar 228 / 200 · +2.6pp PASSED 8/15 early at n=203, re-passed 9/1 (+3.6pp post-recal window) — full weight; re-checked each audit. NOTE: PrizePicks all but stopped offering this market on 8/10, so the counter has barely moved since — the bar is passed, but new evidence is thin
- pitcher-K rung 2 32 / 100 · +4.0pp RE-ENTERED 8/15 at half stakes — rung 2 needs a FRESH 100 at ≥ +3pp for full weight, no rung-1 credit; ≤ 0 re-kills. Window RESET 9/12: it stood at 96/100 (+14.5pp), but all of it was scored by the old model and a new one just started betting — a rung measures the model that is betting, so the count restarts
- 09·12 The $8,443 tuition The whole cumulative P&L curve, every graded pick — a lucky start, a five-week slide to −$8,443, four months grinding back, and a September week that gave it back. 768 picks, net −$615. REPORT
- 08·15 The bet that needed Świątek to lose Report card №3 — the tennis tripwire fires, the autopsy finds the model betting against a floor it couldn't see, and the killed pitcher-K family re-enters at half stakes. REPORT
- 08·01 The first bet my rules refused to kill Report card №2 — batter hits passes the first keep gate, the killed family is beating the market from beyond the grave, and tennis is still five bets from its tripwire. REPORT
- 07·29 My model's worst bet died on schedule Report card №1 — the audit that fired nothing, the kill at exactly n=150, and the bug that ate two weeks of history. REPORT
- 07·28 The loss was never the data noteClaim: the money bleed is a probability-layer problem, not a missing-feature problem. Grades when the recalibration lands. OPEN ⏳
- 07·13 How a losing bet hides inside a winning average Claim: the blend hid a real batter edge and a real pitcher-K hole — and the pre-committed gate would decide, not the mood. It did, five days later. HELD UP ✓
- 07·03 The one number every sports tout hides Claim: closing-line value is the signal that matters, and where this model beats the close the edge is real. Re-grades at n≈500. OPEN ⏳
record so far: 1–0 held up · 2 open rss