The Daily Grind

Monday, June 22, 2026

Underdogs covered the runline 9-3 and won outright 5-7 across 12 graded games (1 postponed). Below: the slate result, the plays our models fired — full stake and reduced (provisional) stake — and the games our discipline guard told us to pass.

Special update — our fourth audit & a new live gameplan

The auto-generated scorecard below says "2 actionable plays, both lost." Taken literally that's true — the only two sides that tripped our strict full-stake rules Monday (Rule B on the Astros and the Angels) both lost. But that one line badly undersells the day, because how we bet just changed. Here is the honest, complete version.

What the fourth external audit found

We periodically hand the whole system — every graded game, the code that grades it, the records it produces — to an outside model (Codex) and ask it to try to break our bookkeeping. The fourth audit (June 22) came back with the core clean: the grader reproduces all 1,148 historical rows except seven we already document, and every season record rebuilds from the raw game log. It also caught four things worth fixing, and we fixed them:

• ROI was hand-counted and had drifted. Our return-on-investment figures were being incremented by hand and had quietly diverged from the data. We now recompute ROI from the game log every night, splitting "real Polymarket price" from "best available price" so neither flatters the other.
• A hidden veto in our pre-game gate. The AI reviewer that signs off on each play had been told to skip near-coinflip underdogs — a filter that isn't in our actual rules. It cost us a winner on June 21 (it skipped the Brewers, who won). Removed: the gate now plays anything the rules qualify unless there's a real conflict.
• Stale metadata and a doc typo on the Rule B band — both corrected.

The uncomfortable finding wasn't a bug, it was the edge itself. Measured honestly — only on games after the rules locked on May 12, and correcting for how many rules we tested — none of our full-stake rules clears the break-even bar with statistical confidence. Rule C (our line-cooling, reduced-stake model) is the strongest of the bunch and still only "promising but marginal" once you pay real prices. We're not going to pretend otherwise: the season-to-date table at the bottom looks great largely because it includes the pre-launch backtest, and the out-of-sample reality (full-stake 17-17) is a coin flip.

The new gameplan: bet everything, tiny, and let the ledger judge

So we changed the goal. Instead of guarding a few "confident" plays, we're now collecting real, forward, money-on-the-line evidence on every promising signal at once — at a stake small enough that being wrong is cheap. Concretely:

• $3.50 a side (Polymarket's 5-share minimum makes a true $1 bet impossible; $3.50 is the floor that clears at any underdog price).
• Every actionable underdog gets both a moneyline and a run-line (+1.5) ticket, so we learn how each signal performs on each market separately.
• We bet not just the old Rule A/B/C fires but also five candidate "edge" signals the audit flagged as worth forward-testing (public-money-fade and away-dog-in-high-totals patterns). Each one is logged with its own tag so its live record stands on its own.
• An AI gate reviews every ticket against the rest of the day's positions and can stand a play down; a hard daily-spend cap and per-game de-duplication keep it bounded.

This is a deliberate, eyes-open override of our own "don't bet unproven signals" rule. The signals are unvalidated — that's the whole point of paying a few dollars to find out in real time rather than guessing. The honest, rules-only record (Rule A/B and Rule C) still lives in the track-record table below, untouched, so you can always see what the disciplined version would have done.

What actually happened Monday: 8-4, +$12.30

Under the new approach we fired six underdogs Monday, twelve tickets in all. Four of the six teams won outright; eight of the twelve tickets cashed.

Eight winning tickets, four losing, +$12.30 at $3.50 a side — our best live day since we started betting real money. Two honest footnotes on the divergence between this table and the strict one below: the Royals were a Rule A play when we bet them in the morning, but by the closing line they no longer graded as a qualified fire — so they win here, in the live ledger, but don't appear in the rules-only record. The Padres were the same story for Rule C: a morning underdog who became the favorite by game time. The bet cashed; the signal record, graded on the close, doesn't claim it. We keep both books precisely so neither can hide the other.

Actionable plays

Actionable plays we fire pre-game: full-confidence (Rule A/B) plays at standard stake. Provisional plays carry a smaller stake while that model is still proving out.

Model track record (season to date)

How to read this: the totals above are season-to-date and include our pre-launch backtest (Mar 27–May 11), when these rules were still being fit to past data — so they overstate what was bettable in real time. The honest out-of-sample record, since the rules locked on May 12: full-stake plays 17-17 (50.0%), reduced-stake plays 71-32 (68.9%). Records are recomputed straight from our logged game-by-game data every night. Stand-down (WATCH) results — 7-4 this season — are excluded entirely, even though they'd improve the headline.

← Back to The Grind