We Published a Wrong Number for Six Weeks. Here's How We Found It and What We Fixed.

On 2026-08-31, we discovered that our settlement grader had been reporting inverted outcomes for run-line bets since 2026-07-22—a scoring bug that turned wins into losses and losses into wins in the published ledger.

What We Got Wrong

On 2026-08-31, we published a daily analysis reporting Sunday 2026-08-30's run-line card as 0-3, -3.00 units. When we verified those same games against MLB box scores, the actual result was 3-0, +1.82 units. The Rangers won 7-4 at the Brewers, the Reds won 7-5 at the Cubs, and the Padres took +1.5 losing 4-5 at the Rays—all three positions won, not lost.

That same day's cumulative book was published as 3-6-0, -4.64 units, -51.56% ROI. The corrected line: 6-3-0, +0.18 units, +2.01% ROI (9 graded).

We publish the losses. We publish every graded bet, win or lose, with the closing line value and the unit outcome. But we also cannot publish fictional losses—and for six weeks, that is exactly what our scoreboard showed.

The Technical Cause

Our settlement grader identified the bet side from the odds sign in the database and applied a home-minus-away margin calculation to determine the win condition. That logic works correctly for moneylines, spreads in other sports, and totals. It does not work for a lane that only ever bets road underdogs at +1.5.

No favorite was ever selected in this lane. All 68 promoted run-line cards took +1.5 on the away team. The grader read a negative odds value, assumed we had bet the home side, and scored the margin backward.

Because both halves of the bug flip results in both directions, the error contaminated outcomes in 86 of 133 graded run-line bets since 2026-07-22.

The corrected lifetime record: 85-48-0, +28.04 units (was 69-64-0, -4.01 units).

Why the Model Still Works

The corrected win rate matches the 63.9% our Phase-2 backtest predicts for this lane's best price bucket, which runs from -180 to -160. The model performed as validated. The scoreboard lied.

The research that built this lane used a holdout sample that the model never saw during training. The backtest win rate was derived before the first live bet was placed. The fact that the corrected live rate lands on the backtested rate, digit for digit, confirms that the scoring method was the variable that broke—not the edge.

What We Corrected and How We Verified It

Every corrected row in the ledger keeps full provenance metadata. A complete pre-repair backup was taken before any write operation. The closing-line-value column, which measures the price improvement between our bet and the market's closing number, was independently re-derived from the raw odds log.

That re-derivation found that 41 of 60 run-line CLV values had been calculated with the same side inversion. The corrected median closing line value is +0.82% (n=59).

The Guards We Added the Same Day

An independent auditor now re-grades the ledger from raw box scores every morning before the report builds. Over the last 14 days, that auditor has checked 379 rows and found zero mismatches.

A regression check pins the exact misgraded case. The volume CSV mirror now self-syncs from Postgres.

What This Means for the Other Live Lane

Hits Props OVER remains our second live lane. Its 7-day closing line value median is -4.21% (n=16), and the lane has lost 3.83 units over the same 7-day window.

Hits Props remains under review. It remains active. The ledger is scored by the same auditor that now re-grades run-line results every morning.

Why We Publish This

We built Signal Syndicate to show the work. That includes the losses. It also includes the times we mis-published them.

The corrected run-line record is better (+16 wins). We are not leading with the improvement. We are leading with the error, because the integrity of the ledger is the only reason the rest of the content has value.

The bug is fixed. The auditor runs every morning. Run-line cards remain active, and every result will continue to be published the day after it grades, with the closing line value and the unit outcome, whether we win or lose.

Related reading

• Closing Line Value
• Signal Score Methodology Memo
• Same Accuracy, Half the Reward — our August 2026 research brief

All figures are estimates. Past analysis is not a guarantee of future results. Not betting advice.