Week 3 stretches across three days — Thursday, Sep 17 through Saturday, Sep 19 — and the slate is a full one: 128 games across the college football landscape. The honest caveat up front is that as of Tuesday, September 15, none of those 128 games carry a posted market line. Zero priced.

That's not unusual this far from kickoff, but it does shape what we can actually say with confidence right now. The model has its internal estimates. The market hasn't moved yet. That gap matters, because the published blend we use leans 70% on the market itself — so until lines post, the edges you'd normally see in a leans table are still forming. Everything tightens as the week progresses and books open their numbers.

The week leads with Thursday night college football, which tends to draw a tighter, more deliberate card than Saturday's sprawling slate. Saturday, Sep 19 is where the bulk of the action falls. If you're tracking the process, the thing to watch is where the market lands relative to our pre-line estimates — that divergence, or the absence of it, is the actual story of Week 3.

The Card hasn't journaled yet — that happens Saturday morning ET, not before. The reason is simple: lines move all week, and any edge identified on Tuesday can look different by Saturday at 11 a.m. Journaling early would mean staking against a number that no longer exists at kickoff.

The mechanism works like this. We run off a fixed Saturday budget — the size stays internal. Each morning we look at the Saturday slate and apply two qualifying floors before anything makes the card: the blended probability on that side has to be at least 55%, and it has to sit at least 5 points above the de-vigged market implied probability. Both floors, not either/or. If a game clears them, it gets sized quarter-Kelly — a fraction of the edge, not the whole thing — and journaled publicly before kickoff. The card can take up to two plays from that process.

With no lines posted yet, we have no qualifying edges to evaluate. That changes as books open their numbers through Thursday and Friday. The card, if it has plays this week, will be posted Saturday morning. If nothing clears both floors, the card is empty and that's the honest entry.

There are no flagged leans this week, and the reason is the same one that shapes the card: zero priced games as of Tuesday. Leans require a posted de-vigged market line to measure against — without that, there's no divergence to track.

That's worth sitting with for a moment. The absence of leans here isn't a quiet week for the model; it's a data availability issue. The model has estimates on all 128 games. But a lean isn't just "the model likes this side" — it's the model's blended probability sitting a meaningful distance from what the market is implying at an executable price. No lines means no gap to measure, and manufacturing one would just be us talking to ourselves.

As lines post through Thursday and into Friday, that picture will update. If anything crosses the published floors — blend side at or above 55%, edge of at least 5 points over the de-vigged market — it'll appear in the tracker with the full context: matchup, day, lean side, blend percentage versus market percentage, and the edge in points. For now, the model and the market are, by default, holding hands. That's the honest read.

The data brief doesn't surface specific ranked matchups or high-total games this week — the top-25 matchups and highest-totals fields are empty, which is consistent with lines not yet being posted. So rather than invent a marquee narrative from thin air, here's what the brief does tell us.

The DFS slate is live, which means the model has processed enough game-environment information to generate player projections across 125 players. That's useful context: when projections are running, the underlying game reads exist even if market prices don't. Ohio State's Jeremiah Smith leading the slate at a 23.8-point projection suggests a game environment the model finds favorable for volume. Duke's Nate Sheppard at 21.7 points projected from the running back position points toward a game the model expects to run through that backfield. Georgia carries Gunner Stockton as the top quarterback projection at 19.6 points, with a ceiling that reaches 52 — the highest ceiling on the board.

Those are the games worth watching as lines post. The matchup context behind each projection will sharpen considerably once spreads and totals are available, and that's when the leans tracker will have something real to say. For now, the DFS section below is where the week's most concrete forward-looking numbers live.

Walk-forward validation means we replayed the model from preseason priors only — no peeking at outcomes — across the 2021 through 2025 seasons. The question was whether it could beat naive alternatives a reasonable person might use instead. It did, on two tasks.

On game-by-game probability (measured by log-loss, where lower is better): the model's blended output scored 0.5684, compared to 0.6931 for a flat 50/50 guess, 0.6603 for a home-team climatology model, and 0.5779 for a no-margin Elo. On season win totals (measured by mean absolute error, where lower is better): the model's 1.959 beat using last year's win total as a forecast (2.188) and flat .500 (2.27).

Those are real improvements over credible baselines, and they're measured on data the model never saw during construction. That matters. What also matters is the honest limit: validated against baselines is not the same as validated against the betting market. The published blend leans 70% on the market itself, which is an acknowledgment that the market is an exceptionally good forecaster. We're not claiming to consistently beat it.

The in-season record through Week 3: 433 games graded. The blend has logged a 0.3643 log-loss and 84.55% accuracy across 123 graded games. The market, over the same 123, is at 0.3444 log-loss and 87.8% accuracy. The market is outperforming the blend on this sample. No guard note is active. Small samples are noise — this is a small sample.

The main slate is live with 125 players. These projections are base offensive-point estimates built from projected passing, rushing, and receiving opportunities, opponent context, and positional priors. They exclude fumbles lost, two-point conversions, return touchdowns, and offensive fumble-recovery touchdowns — historical full-score validation is paused until those outcomes are reconciled. The uncertainty ranges are experimental. With that said, here's where the model sees value across the top of the board.

• Jeremiah Smith, WR, Ohio State — $8,400 salary, 23.8-point projection, 2.83 value (proj/salary per $1,000). Floor of 8, ceiling of 42.5. The highest value score among the top names, at a salary that isn't at the very top of the position. The wide range between floor and ceiling is worth noting for cash versus tournament decisions.
• Nate Sheppard, RB, Duke — $9,200 salary, 21.7-point projection, 2.36 value. The highest ceiling on the running back side at 45.1, though the floor of 1.2 is essentially zero. GPP-relevant range.
• Gunner Stockton, QB, Georgia — $9,200 salary, 19.6-point projection, 2.13 value. The highest ceiling on the board at 52.0, and a floor of zero. Classic high-variance tournament quarterback.
• Marcel Reed, QB, Texas A&M — $8,700 salary, 18.3-point projection, 2.10 value. Ceiling of 41.4. The lower salary relative to Stockton gives him a slight value edge.
• Reed Harris, WR, Arizona State — $7,300 salary, 17.5-point projection, 2.40 value. The most affordable name among the top six with a ceiling of 42.4 — that combination of lower salary and competitive ceiling makes him worth considering in both formats.
• Ashton Daniels, QB, Florida State — $6,400 salary, 17.2-point projection, 2.69 value. The second-highest value score on the board after Smith. At $6,400 he's the cheapest quarterback here by a meaningful margin, which creates salary flexibility elsewhere in a lineup.

Low ownership is not inherently leverage, and projected ownership numbers aren't in this brief — so that's not a factor we can speak to. The game environment behind each of these projections will come into sharper focus as lines post. Build lineups with the floors and ceilings in mind, not just the means.

Week 3 is a full slate — 128 games across Thursday through Saturday — and right now it's a preview without prices, which means the most honest thing we can do is explain the process and wait. The card journals Saturday morning against whatever lines are actually posted; the leans tracker populates when the model finds a gap worth flagging against the de-vigged market. Neither happened yet, and we're not going to fill that space with manufactured conviction. The DFS projections are the most concrete forward-looking numbers available today, and they carry their own caveats about scope and experimental ranges. The record is public — blend at 84.55% accuracy over 123 graded games, the market outperforming at 87.8% on the same sample. Both the card and the leans grade in public, the misses exactly like the hits. When lines post, the week gets real.