Week 8: what our 10–4 record actually tells us

The result needs a baseline

The winner record was 71.4%. Week 4 had more correct picks, eleven, but sixteen games gave it a lower winner rate of 68.8%. Neither result describes every type of prediction: Week 8 spreads finished seven and seven, while totals finished six and eight.

We compared the model with the favorite implied by the preserved pregame spread on the identical fourteen games. Both methods won ten. They agreed on twelve selections. Choosing Denver over Kansas City gained one correct winner; choosing Minnesota over Detroit lost one. The two disagreements offset each other, leaving no net winner advantage over that baseline this week.

This comparison is about winner selection, not betting return. A favorite's moneyline price matters for profitability, and ten correct winners alone do not establish a profitable strategy.

Every selection in the sample

These are the original recorded picks and final scores, not forecasts rebuilt after the results. Each matchup links to its evidence view. The downloadable record includes the locked lines, timestamp, version identity, model/input hashes and calculation rules.

Download the frozen Week 8 results and methods (JSON)

Four misses, four different questions

Dallas: a large estimate with incomplete evidence

The saved forecast selected Dallas at 84.3%, with a home margin of +15.7 points. The locked market expectation was Dallas +7. Arizona won 33–30 in overtime, making the realized home margin −3. The forecast missed that margin by 18.7 points.

The strongest concern is not simply that an upset occurred. A genuine 84% event can lose. The issue to investigate is whether the model's large expected advantage and narrow-looking winner estimate are supported when personnel evidence is incomplete. Both team availability contexts were marked partial. The oldest accepted report ages were approximately 6.9 days for Dallas and 5.8 for Arizona at forecast time. Those ages do not identify a specific missing injury or prove why Dallas lost.

The saved explanation also incorrectly presented line movement as confirming Dallas. The retained Dallas spread moved from −10.5 to −7, toward Arizona, across roughly sixty days. We corrected that display contradiction and its invented point attribution without altering the original pick. That price history does not prove rapid steam or identify professional bettors.

Seattle: investigate the size of the error

Seattle was selected at 66.6%, with a projected home margin of +6.3. Chicago won 34–17, a home margin of −17: a 23.3-point error. The stored Chicago availability context was about 10.1 days old. Reviewing role coverage and football assumptions is warranted; calling this only a close-game swing would miss the size of the forecast error.

Minnesota and the Jets: small estimated advantages

Minnesota's 53.9% selection lost 19–15 to Detroit. The Jets' 50.6% selection lost 38–33 to Las Vegas. These were near-even estimates. Their losing outcomes should not be presented as the same kind of evidence as a high-probability miss. The saved football-only lane narrowly favored Las Vegas in the Jets game; switching to that lane retrospectively would improve this week's record, but choosing a rule after seeing fourteen outcomes is not a valid promotion test.

Do the probability groups separate outcomes?

Selections displaying 70% or more went seven and one across eight games. That is an encouraging description of this sample, not proof that a 70% filter improves future picks. Thresholds chosen from a single successful week risk fitting noise.

The sum of displayed winner probabilities implied 9.94 expected wins, close to the ten observed. Aggregate agreement can hide individual errors. The winner Brier score, the mean squared difference between each selected team's probability and its binary result, was 0.1848. A lower score is better. This fourteen-game value does not establish calibrated probabilities for the current engine.

Score quality deserves its own review

The projected home-margin mean absolute error was 11.78 points. The locked spread's implied home margin had a lower error of 11.11. Matching the favorite's winner record therefore did not mean matching its margin accuracy.

The totals record of six and eight does not justify increasing or decreasing every projection by a fixed amount. Pace, possessions, scoring opportunities and team/player workload need coherent treatment. A change should be tested on held-out games and future paired forecasts, with separate limits for winner, spread, total and player-prop performance.

How this review was built—and its limits

The evidence cutoff is the set of saved 2026 regular-season Week 8 selections and settled outcomes inspected November 3. All fourteen use the one-hour pregame lock policy. Winner grades were checked against final scores; spread grades use the selected team's margin plus its locked spread; totals use the final score sum, locked line and original over/under direction. There were no ties or pushes in these fourteen graded selections.

Each original snapshot was matched to its retained version and input/model identifiers. No current roster, new sportsbook quote or postgame model recalculation was substituted. Report ages describe the evidence at forecast time, not its age when you read this article. Results in this frozen download remain fixed even as the live season scorecard grows.

The earlier season includes five-minute locks and different model-artifact vintages. Across 120 decided historical games, the saved model went 71–49 versus 76–44 for the locked-spread favorite baseline; one tied game is excluded. This mixed history is not a clean test of today's engine. Profile and explanation improvements released afterward cannot receive credit for producing these past outcomes.

All twenty-eight team availability contexts in Week 8 were partial. Neither their incompleteness nor an individual upset establishes causation. Next priorities are dated quarterback/role/replacement evidence, coherent opportunities and empirically tested calibration. We have not promoted a new numerical rule based on this review.

Sources, reproducibility and corrections

• Frozen original selections, locked lines, final outcomes and methods
• Live scorecard and same-game favorite comparison · Public Week 8 accuracy data
• Arizona–Dallas saved evidence · Retained forecast versions
• Definitions, data sources and limitations
• AI assistance and editorial disclosure · Report a public correction

Published by Pick6Labs. This article was assembled with AI assistance; its numerical claims were checked against the retained records and reproducible calculations. It is an original site-results analysis, not independent peer review, film scouting, a gambling guarantee or a claim of future profitability.