NFL game totals · 2026 season · picks with receipts

ORACLESee the edge.

A machine-learning model for NFL game totals. Every pick posted here before kickoff, every result graded in public, and it tells you when it's cold.

The ledger

Week 1.
On the record.

This is the product. Each pick is stamped at the line it was released at, and stays here graded win or loss. The ledger is append-only: rows are added, never edited.

Board priced Sep 3, 2026 · 23:05 UTCSeason 2026 · Week 1Status OPEN
New England at SeattleWeek 1 · NE @ SEA · released Sep 3 · 18:05 UTC
UNDER 44.5
39.5
PENDING
Denver at Kansas CityWeek 1 · DEN @ KC · released Aug 30 · 08:20 UTC
UNDER 43.5
34.9
PENDING
14

games the model priced and passed on. It needs a projected edge of at least 5 points before it releases a pick. Most weeks that means a short list. Some weeks it means none, and we say so.

Ethan, who built Oracle, sitting outside at night

What we're doing here

I'm 22. I built a model. Now I'm going to be wrong in public.

Every sports pick account you've seen sells the same thing: a screenshot of a win and a promise. Oracle sells the opposite. A model, a ledger, and a rule that it has to prove itself before anyone pays for it.

The model prices every NFL game total from twenty inputs: how efficiently each offense has been moving the ball, recent scoring against its own baseline, wind and cloud cover at kickoff, quarterback quality, short rest, primetime and divisional flags, and how lopsided the spread is. It only plays UNDERs, because that's the only side where the backtest held up honestly. When it doesn't see an edge, it sits. When it goes cold, it halts, and the halt goes on this page too.

You won't find a win-rate on this site. Not because it's bad, but because a percentage on a landing page is worth nothing until it's earned live. What you'll find is every pick, timestamped, and what happened next.

— Ethan, builder of Oracle

How it works

Three steps.
No black box.

The whole process fits in a paragraph, and it doesn't change week to week. That's the point.

01

Price every game

Tuesday through Saturday the model projects a total for all sixteen games from twenty features and compares it to the market line.

02

Release only the edge

A pick goes out only when the projection sits 5+ points under the line. It is posted here and to the feed before the game starts, with the line it was priced at.

03

Grade it, then grade the grader

After the game the pick is marked win or loss. We also record the closing line and show each pick against it, because beating the close is the one number that separates skill from luck.

The bar, written down before Week 1

Nobody pays until the closing line says the model is real.

Every pick is a public paper trade until at least 30 games have been graded against the closing line, the minimum sample we wrote down before the season. The billing switch stays off until that sample clears the bar. If it misses, this page will say it missed.

≥ +0.5 pts
≥ 60%
0 / 30

Plans · after the gate

Pro-grade model.
Lunch-money price.

Apps like this charge $30 to $100 a month. Oracle starts at $10, because we'd rather prove it first. Billing opens only if the bar above is cleared.

Starter

$10/ MO

Up to 3 picks a week

  • Picks before kickoff
  • Full public ledger
  • Halt-week alerts
Most useful

Pro

$20/ MO

Up to 10 picks a week

  • Everything in Starter
  • Model notes on every pick
  • Closing-line report weekly

Elite

$40/ MO

Every pick the model releases

  • Everything in Pro
  • Some weeks that's 3, some weeks 10
  • First access as sports are added

7-day free trial at full Elite access when billing opens. Founding-member price lock: early subscribers keep $10 / $20 / $40 for life, even as sports are added and prices rise. No lifetime deals, no annual plans.

Follow the season

Watch it happen
before it happens.

This page is the record. The feed carries the weekly ledger card, the picks, and the halt weeks. The app arrives around Week 9. Until then, this page and the feed are the product.