Cleveland Won 27–24. Right Side, Low Total

Cleveland beat Pittsburgh 27–24, and both forecast versions got the winner right. The rebuilt point model projected a 3.2-point Browns margin—close to the actual three—but its 43.9-point total came in 7.1 points low.

The Thursday preview put Cleveland at 70.4% in the original final forecast and 51.9% in the rebuilt research forecast. It also preserved Cleveland +2.5 and Over 38.5. All three selections won. Friday’s report supplies the result and grades against those unchanged pregame forecasts.

That gives us several things to separate: picking the winner, estimating how close the game would be, and anticipating how much scoring it would produce.

Cleveland built a lead, then needed one more drive

Cleveland scored all three of its touchdowns in the second quarter and led 21–10 at halftime. That remained the score entering the fourth. Jaleel McLaughlin’s 28-yard touchdown followed a Denzel Ward interception; Quinshon Judkins and Harold Fannin Jr. supplied the other touchdowns. The Browns’ recap records the burst that put Pittsburgh behind.

The Steelers came back. Aaron Rodgers’ touchdown pass to Pat Freiermuth and ensuing two-point sneak tied it 24–24 with 1:42 left. Deshaun Watson moved Cleveland into field-goal range, and Andre Szmyt hit from 56 yards with ten seconds remaining. Cleveland improved to 3–1, its third straight win; Pittsburgh fell to 2–2. Steelers recap

Watson finished 24 of 33 for 268 yards, one touchdown and one interception. Rodgers went 22 of 40 for 299 yards, three touchdowns and two interceptions. Cleveland held a narrow 372–361 total-yardage advantage, and both teams committed two turnovers. The protection concern from the preview did not become a decisive Pittsburgh advantage: Cleveland allowed two sacks, while the Steelers allowed five. Those figures come from the official gamebook.

This was a game Cleveland had to finish after surrendering its lead. The final margin describes that closeness; it does not tell us which pregame win probability was properly calibrated.

The predicted margin held up better than the points

The saved pregame files contain point projections that Thursday’s article did not display. These come from the XGBoost total and margin models, with team scores calculated as (total ± margin) / 2. They are distinct from the five-model winner blend and the spread/total probability aggregates.

The rebuild missed the margin by just 0.19 points, compared with 1.01 points for the original. Its implied scores were both too low: Cleveland by 3.45 points, Pittsburgh by 3.64. Similar scoring errors largely canceled when taking the difference between the teams.

The original total missed low by 8.09 points; the rebuilt total missed low by 7.09. Calling the winner and nearly matching the margin did not mean predicting the scoring level accurately.

These projections imply Cleveland as roughly a two-point original favorite or 3.2-point rebuilt favorite. The saved bookmaker handicap was Cleveland +2.5, making Pittsburgh the market favorite. A model’s predicted margin and the handicap used to grade a cover are different quantities.

Winner, spread and total all landed on the selected side

Here are the saved aggregate probabilities for each winning selection. The original market aggregates retain their older policy, including T = 0.4 sharpening; the rebuilt spread estimate combines three trained models, and its totals estimate also pulls the result toward 50%. These are different forecast policies.

The cover calculation is 27 − 24 + 2.5 = 5.5. The total clears its saved threshold by 51 − 38.5 = 12.5. Thursday afternoon’s market median had fallen to 37.5, but the preserved totals probability and recorded grade remain attached to 38.5.

Among the trained spread models, all three original picks hit. In the rebuild, XGBoost and Random Forest hit while Logistic’s Pittsburgh -2.5 pick missed. For totals, the original linear model’s slight Under lean missed; its other two models hit. All three rebuilt totals models selected Over and hit.

Each aggregate starts Week 4 1–0 in its own market. These are three related outcomes from one game, not three independent wins establishing a betting advantage or a record of wagers placed.

The stronger original confidence earns the better winner grade

Four of five winner models selected Cleveland in both versions. Elo alone picked Pittsburgh at 60.7%, and missed. The table grades every contributor and both ensemble stages. Each cell is Brier score / natural-log loss; lower is better.

With Cleveland winning, the calculations are (1 − p)² and −ln(p), using each forecast’s full-precision Cleveland probability. The original raw blend was 58.6%; sharpening raised it to 70.4% and improved this game’s Brier score by 0.0840 and log loss by 0.1840.

The rebuilt forecast’s T = 1 leaves raw and final identical. It receives the larger penalty because it assigned Cleveland less probability. A three-point win receives exactly the same winner-probability grades as a twenty-point win would.

After sharpening amplified the Atlanta miss, this result rewards it. Both belong in the record. One Cleveland win cannot validate sharpening, prove the rebuild worse, or establish calibration. The rebuild still uses the raw fallback after its proposed calibration failed independent checks.

Sunday brings the disagreements

The first test is Indianapolis against Washington in London, Sunday at 8:30 a.m. Central, confirmed in Washington’s preview. Both versions take Washington: 77.3% original, 54.4% rebuilt. Friday’s 12:42 p.m. Central snapshot puts Washington’s median paired, margin-removed market probability at 35.7% across ten books. That is a substantial disagreement to follow, not a certified edge.

Then come the three changed winners, all scheduled for noon Central: Giants to Arizona, Dallas to Houston, and Tampa Bay to Green Bay. The rebuilt picks are only 53.4%, 52.4% and 52.9%, respectively. Those games can separate the versions on winner accuracy; Cleveland could only separate their confidence scores.

The full Week 4 comparison keeps both boards visible. Fifteen games remain pending: fourteen Sunday and one Monday. Sunday’s results will test the changed picks and the reduced confidence together, against the forecasts already saved.

Source note: Game 2026_04_PIT_CLE finished CLE 27, PIT 24 on October 1. This review uses Friday’s report at prediction-repository revision 6bfc6ef, the preserved original forecast, the pregame research moneyline freeze and separate market forecast, and the official sources above. Companion editorial evidence retains point-model lineage, full-precision probabilities, grading calculations and source hashes. Week 4 research tracking remains separate from the older 48-game performance monitor. Writing this article did not refit models or change forecasts, grading records or the betting ledger.