NFL Thursday Special — SF @ LAR (Melbourne)

The first NFL regular-season game ever played in Australia is a single-game slate, and the honest read is that there's nothing to stake. The model and the market agree on direction — Los Angeles is the side — but the total is where they split, and even that gap isn't clean enough to bet. A new player-sim layer adds the "who scores" texture without moving the total.

Data Freshness

• Venue: Melbourne Cricket Ground (MCG), Melbourne — neutral site, SoFi not in play
• Kickoff: Thu Sep 10, 10:35 AM AEST (7:35 PM ET)
• Surface: Matrixturf (artificial), open-air
• Weather: clear, cool, calm — ~53°F at kickoff, winds 2–4 km/h (dead calm), 0% rain. A genuine non-factor.
• Market: 9-book decision snapshot, Sep 8 23:25 UTC
• Model: epa-ratings-v2, frozen Sep 10 12:20 UTC (SHADOW tier)
• Sim: 50,000 draws, SHADOW_UNCALIBRATED (scenario tool, cannot authorize a stake)
• Player sim: player-stats-shadow-v1 (NEW, 50k runs) — SHADOW_UNCALIBRATED, scenario only
• Rosters: 2026 depth verified via web only — Bosa, Kittle, Garrett, Donald all expected to play with rust risk

Slate Overview

One game, three layers that mostly agree. The model leans LAR -4.8 with a 46.2 total; the market sits at LAR -3.5/-4.0 with a 48–48.5 total; the sim gives LAR a 62.8% win probability and leans under. The spread is a dead zone. The total is the only number with a gap — and it's a watch, not a stake. The new player layer adds texture: CMC and Nacua are the two names to watch, but everything stays shadow-tier.

Game-by-Game Breakdown

SF @ LAR (neutral — MCG, Melbourne)

The model and market agree Los Angeles is the side, but the spread offers nothing. LAR -3.5 covers 53.9% in the sim against a fair -117, and the best market price is -112 — a point-and-a-half of edge that evaporates once you account for the fact that this model is shadow-tier and uncalibrated. San Francisco +3.5 is worse: 46.1% against a fair +117, no value at -108.

The total is the real story. The model projects 46.2, the sim puts the under 48.5 at 58.6% (fair -142), and the under 48.0 at 58.7% against a LowVig -104. That's the widest gap on the board — two aligned model-side signals pointing under.

But here's the counterweight, and it's why this stays watch-only. The epa-ratings-v2 model failed its benchmark, and totals are its least reliable output. The 2025 head-to-head went 49 and 68 — two games that blew past this number. And the market is drifting the other way: BetRivers moved 48.5 to 49, mybookieag and bovada ticked the spread up. When the model and the market momentum point in opposite directions, the honest call is to downgrade confidence, not force a position.

The travel asymmetry is a wash. San Francisco landed a full week early and practiced on-site; Los Angeles flies in the day before and plays on a Pacific body clock (kickoff is 5:35 PM PT, inside the performance window). The two effects partially offset — net noise, not an edge.

The player layer — who scores, not whether the total moves

The new player-sim layer (player-stats-shadow-v1) allocates the team sim down to individual players, and it surfaces two names worth watching.

Christian McCaffrey is the single most interesting number on the board: 46.8% anytime touchdown, the highest on the slate. Against a 0.90 team rush-TD baseline, the sim sees him as nearly half the San Francisco end-zone pie — 94 rush yards [68–123] and 21 receptions. That's a real shape to watch against a live ATTD price.

Puka Nacua is the receiving market's highest-variance name: 72 rec yards [45–108] on 7.4 targets, the widest IQR of any receiver on the board. That 45–108 band is the story — a point estimate would hide it.

The rest of the board: Purdy 239 pass yards [197–284] with 0.94 INT, Stafford 269 [214–326] with 0.88 INT. Kittle 53 rec yards [32–77] — but that number carries an asterisk the sim can't see: he's under nine months post-Achilles, first real game back, and the sim's last-6 weighting has no rust adjustment. Treat his IQR as optimistic until gameday status confirms.

The reconciliation guarantee is the tell of a serious build: on every simulated path, QB attempts equal team attempts, QB completions equal the sum of receiver catches, QB pass yards equal the sum of receiving yards, and team carries equal the sum of rushes. The "Team Other" bucket makes unassigned volume visible rather than hidden — SF carries a 16-yard residual, LAR about 17, the rookies and no-history signings the sim can't place.

Targets: Under 48.0 (LowVig -104) is watch-only. Spread pass both sides. CMC ATTD and Nacua rec yards are shadow-only shapes to watch against live prices, not stakes.

Show Your Work

Under 48.0 (LowVig -104) — watch, no stake

• Model: epa-ratings-v2 total 46.2 (SHADOW tier). Sim under 48.0 at 58.7%, fair -142 vs -104.
• Confidence band: 52–60% under. The gap is real on paper, but the model is uncalibrated and totals are its weakest output.
• Assumptions: The sim's distribution holds despite the shadow-tier flag, and the 2025 H2H totals (49, 68) are not predictive of a neutral-site opener.
• Key variables: Whether the market keeps drifting up (BetRivers already at 49), and whether the calm weather holds — it should, which removes the one environmental reason to lean over.
• Grade it by: If the game lands under 48, the read was directionally right but un-staked — a process win, not a cash win. If it lands over, the market drift was the signal and the model's total was the noise.

CMC anytime touchdown — watch, no stake

• Model: player-stats-shadow-v1, 46.8% any-TD, 94 rush yds [68–123], 21 rec.
• Confidence band: 40–52% any-TD. The shape is real, but the sim has no snaps/routes and is uncalibrated.
• Assumptions: CMC's usage holds at his recent share, and the sim's last-6 weighting isn't distorted by his 2025 workload.
• Key variables: The live ATTD price, and whether the official inactive list (not roster ACT) confirms him active ~90 min out.
• Grade it by: If CMC scores, the 46.8% was directionally right. If he doesn't, the sim's end-zone share was the overreach — and the calibration gap is the data point that eventually promotes the tier.

Slate Shortlist

• Watch (not stake): Under 48.0 LowVig -104 — model + sim aligned, but SHADOW_UNCALIBRATED and market drifting opposite.
• Watch (shadow): CMC anytime TD (46.8% sim) and Nacua rec yards (72 [45–108]) — shapes to compare against live prices, not stakes.
• Pass: LAR -3.5 everywhere (fair -117, best market -110 — no gap), all props (usage model unusable).

Notes Before Lock

• Sim and player-sim are SHADOW_UNCALIBRATED — neither can promote a tier or authorize a stake.
• Official inactives are the gate for all player markets — Bosa and Kittle status verify ~90 min out.
• No decision-time snapshot yet; the Sep 8 capture is the latest on file. Decision-time lines required before any stake.
• Weather is clear — no environmental downgrade trigger.
• QB status confirmed: Purdy (SF) vs Stafford (LAR), both expected starters.
• Bosa (ACL), Kittle (Achilles), Garrett (knee), Donald all expected to play — rust risk, not a scratch.

TL;DR

• Under 48.0 (LowVig -104) — the only number with a gap, and it's watch-only: uncalibrated model + market drifting the other way
• CMC anytime TD — 46.8% sim, highest on the board, shadow-only
• Nacua rec yards — 72 [45–108], widest IQR, shadow-only
• LAR -3.5 — pass, no stakable edge after shadow-tier noise
• SF +3.5 — pass, no value at -108
• Weather — non-factor, calm and clear at the MCG

Updates

• Revision 2 (11:00 AM EDT, September 10): Added player-sim layer (player-stats-shadow-v1, 50k runs). New player distributions: CMC 46.8% any-TD (highest on board), Nacua 72 rec yds [45–108] (widest IQR), Purdy 239 / Stafford 269 pass yds, Kittle 53 [32–77] with post-Achilles asterisk. Under-lean unchanged. All player numbers shadow-only, no stake change.
• Revision 1 (9:30 AM EDT, September 10): Initial Thursday special. Single-game slate at the MCG. Model LAR -4.8 / 46.2 (SHADOW), market -3.5/-4.0 / 48–48.5, sim LAR 62.8% / under 58.6%. No stakable play — under 48.0 is watch-only, spread and props pass. Weather a non-factor.