College football's opening stretch ran from August 29 through September 5 and produced 301 games on the schedule. Of those, 295 graded upstream — 6 remain unresolved and will stay open in the journal until results come through. We're not filling those gaps with guesswork.

The blend model processed 111 games it had probability estimates for, and the broader season record now sits at 433 graded results. That's enough to start reading the shape of things, not enough to draw conclusions about edges.

The games that defined the week weren't the ones we expected. TCU Horned Frogs — a 72–73% blend favorite depending on the snapshot — fell to North Carolina Tar Heels, 15–10. UNLV Rebels, at 65% in the blend, lost to Memphis Tigers, 27–21. And on Thursday night, Georgia Tech Yellow Jackets came in as a 71% blend favorite and dropped a 14–13 decision to Colorado Buffaloes. None of those were coin-flip games in the model. All three went the wrong way. That's what upsets are — not anomalies that break the model, but the 28–35% outcomes that the model said would happen sometimes. The question worth sitting with is whether they happened for reasons the blend missed, or whether 301 games into a season is just too early to know.

There is no card to grade this week. No play cleared both edge floors before the Saturday slate, so nothing was journaled and nothing was staked.

An empty card is a position — sitting out when the numbers don't warrant action is the same discipline as playing when they do. The card record remains blank, and that's the accurate entry.

The tracked leans — informational, unst­aked, graded publicly — went 6-0 this week. Here's each one by name.

• Eastern Michigan Eagles vs. Sacramento State Hornets (Saturday, Aug 29): blend 81%, market 73%, edge 7.4 percentage points. Eastern Michigan won 28–17. Hit. (This matchup appears twice in the journal with a trivial total-line difference — 53 and 53.3 — both rows grade the same result.)
• James Madison Dukes vs. Liberty Flames (Saturday, Sep 5): blend 73%, market 67%, edge 5.6 pp (5.5 pp in the second journal row — same game, same result). James Madison held on, 20–13, as a 6.5-point home favorite. Hit. The final was tighter than the spread implied; worth noting.
• Northwestern Wildcats vs. South Dakota State Jackrabbits (Saturday, Sep 5): blend 88%, market 82%, edge 5.6 pp. Northwestern won 34–18 as a 13.5-point favorite and covered comfortably. Hit.
• Temple Owls vs. Rhode Island Rams (Saturday, Sep 5): blend 84%, market 78%, edge 5.3 pp. Temple won 38–14 against a 54.5 total market and covered the 11-point spread with room to spare. Hit.

6-0 is a good-looking number. It is also six games. The lean record will be more meaningful at 60, and the upsets in the same week are a useful reminder that the blend was wrong about TCU, UNLV, and Georgia Tech in roughly the same breath. Own both parts of that sentence equally.

Season-to-date, 433 games have graded across all tracked results. The week's blend evaluation — the 111-game subset with model coverage — came in at a log loss of 0.3297 and accuracy of 84.68%.

The fuller season record (123-game subset) shows blend log loss at 0.3643, Brier score 0.1133, and accuracy of 84.55%. The market over the same 123 games: log loss 0.3444, Brier 0.1075, accuracy 87.80%.

The market is outperforming the blend on every calibration metric right now. Log loss and Brier score both measure how well-calibrated the probabilities are — lower is better — and the market is sharper on both. The blend's accuracy gap versus the market (84.55% vs. 87.80%) is meaningful in direction, even if 123 games is a thin sample for confidence. We've been direct about this from the start: the blend has not proven itself against the market out-of-sample, and one week of lean results doesn't change that accounting.

The leans performed well. The model-vs.-market calibration gap is real and worth watching. Those two facts can both be true at the same time, and neither one cancels out the other.

The TCU, UNLV, and Georgia Tech results are the more instructive data. Three games where the blend had 65–73% confidence, all three going the other way. That's not a crisis — the model said it would be wrong roughly that often — but it's a reason to look at what the blend was weighting in those matchups before Week 2 results pile on top.

Six unresolved journal rows will close when upstream data catches up. The card floors stayed where they are. If an edge clears both thresholds before Saturday, there will be a journaled play; if not, another empty card week is the right call. The leans will keep tracking either way — informational, public, graded honestly when the final scores come in.