The engine, without the smoke machine.

What Atlas does

Separate research agents inspect official player availability and roles, venue weather, available sportsbook quotes and opposing analysis. Atlas resolves conflicts, writes the forecast and publishes a structured evidence record. Operator articles can supply a quoted price; they are not treated as an independent model verdict.

Our winner and five player picks

Each new matchup report names a predicted winner, projected score and at least five distinct player outcomes. Yardage targets are our own at-least thresholds, rather than copied sportsbook lines. A 60+ receiving-yard forecast succeeds at 60 yards. A touchdown pick counts the player scoring, not throwing a passing touchdown. Sleeper labels identify less established or more volatile opportunities; they do not mean guaranteed betting value.

What the percentages mean

A 60% estimate means we judge that specific outcome more likely than not. These are subjective research estimates, rounded to five percentage points, with source-backed role and statistical anchors and explicit scenario assumptions. They are uncalibrated: we have not yet demonstrated that events labeled 60% actually happen 60% of the time. A season average is not an event hit rate. Player forecasts assume the described availability and role; material changes require a new pregame revision. Correlated picks are not a suggested parlay.

Earlier editorial v1

A watchlist identifies a player who merits more research. A call takes a side on a specific outcome. A yardage projection is an editorial scenario with its arithmetic and assumptions disclosed. None is a calibrated probability or a guarantee of betting value.

Source priority and freshness

Official injury and inactive reports take precedence over previews. A blank injury page does not mean healthy. A weather forecast is not an observation. An operator-published quote is not necessarily executable. Movement requires two timestamped prices from a comparable market; otherwise it is marked not measured.

The permanent record

Each publish creates an immutable revision and source snapshot. Edits point to the prior revision. The database rejects new game calls at kickoff; a verified pregame scan can lock them earlier. Article content, its timestamp, quoted lines and assumptions remain attached. An administrator could still deploy new database code, so this is an application audit trail, not an independently notarized ledger.

What learning means

After verified final results, the system grades only the final eligible calls. It reports sample size, win/loss/push/void and yardage mean absolute error. Missing results stay ungraded. The weekly review identifies faulty assumptions, proposes versioned experiments and compares them across held-out weeks. One ugly Sunday is not enough evidence to rewrite the method.

Research v2 grades winner and individual player probability error with Brier scores, alongside hit rates, sample sizes and yardage error. Calibration requires enough completed forecasts across probability bands and held-out evaluation; reporting a Brier score alone does not establish it. Earlier editorial v1 published no probabilities, and those historical values remain empty. It claims no ROI from indicative prices and has no verified placed-bet ledger. This process updates the product’s evidence and methodology; it does not automatically retrain the underlying AI or ensure improvement.

Public work, private source notes

Public pages show factual citations and our reasoning. Expert identities and raw counterarguments stay in the owner-only source room. New signups go to EmailOctopus for email confirmation. EmailOctopus manages subscriber preferences and unsubscribe links. Connecting signup does not by itself schedule daily newsletters.