138 games. 138 graded. A clean sheet on the administrative side, which is the boring kind of clean sheet — the interesting one belongs to Oklahoma State.

The Oklahoma State Cowboys beat the Oregon Ducks 39–31 on Saturday. That result defined Week 2 for the simple reason that Oregon was an 85–86% blend favorite — depending on which snapshot you pull, both appear in the data — and Oklahoma State beat them by eight. Not a fluke-field-goal margin, not a lucky bounce at the end. Eight points. On the scoreboard.

A result like that deserves to be described plainly before it gets explained. Oregon was a heavy favorite. Oklahoma State won. That's the week.

The rest of the slate was largely chalk. The model had 12 games in the blended sample for the week and went 10-for-12 on winners — an 83.3% accuracy figure that sounds good until you notice the one miss that matters most was the biggest number on the board. That's how upsets work: they're rare, so the accuracy stays high, and the logloss absorbs the damage quietly. We'll come back to that in the record section.

No card was journaled for this week's Saturday slate. No edge cleared both floors before kickoff, so nothing was staked — and an empty card week is a position in itself. The record doesn't change; neither does the accounting.

We will not retrofit a winner to the card after the fact.

No leans were flagged this week either. Where the model had a view, the market largely agreed — and when the agreement is tight enough, there's nothing useful to track publicly. A week with nothing flagged isn't a failure of effort; it's the filter doing its job.

Lean record for the week: 0–0. Season lean record: 0–0.

Through 433 graded game-sides on the season, the blended model has an accuracy of 84.6% against a market baseline of 87.8% across the 123-game window where both are tracked in parallel. Blend logloss is 0.3643; market logloss is 0.3444. The market is outperforming the model on both metrics at this sample size.

That's an honest read of a small-sample, early-season record. 123 games is not enough to draw firm conclusions about which approach is better calibrated over time — the brief carries no guard note this week, but the gap between blend and market is real and worth watching. The model leans on the market at a documented weight for exactly this reason: early-season calibration data is thin and the market processes information we don't have.

The Oklahoma State–Oregon result is a useful stress test of how logloss works. A single 85–86% favorite losing doesn't move accuracy much in a 12-game sample — going 10-for-12 is still 83.3% — but it does move logloss, because the model assigned a low probability to the outcome that happened. That's appropriate penalty for a confident wrong call, and it should be.

The Oklahoma State result is a reminder that an 85% probability means something loses roughly one time in six. The model isn't broken when that happens; it's behaving as advertised. What matters is whether the 85% calls are well-calibrated across a large enough sample — and 123 games is not yet that sample.

The market is currently outperforming the blend on both accuracy and logloss. We're not going to paper over that with early-season noise caveats; it's in the record and it's the right thing to track. The blend leans on the market intentionally, which is the right structural response to this phase of the season, but it hasn't yet justified the weight it adds on top of the market signal.

Nothing goes on the card or the lean tracker until edges clear both floors. The quiet weeks are part of the process — a position with nothing staked is still a position. Week 3 will tell us more.