An expected-value number is only as honest as the price underneath it. If a feed says one matchup ranges from a home favorite to a home underdog, the answer is not to average the endpoints and print another decimal. The answer is to stop and inspect the market rows.
The August 21 packet is a useful stress test because it contains both tidy-looking aggregates and implausibly wide ones. It does not prove an EV tool was rebuilt, deployed, or guarded by any particular threshold. This column is the input contract such a tool should satisfy: show the exact offer, expose the range, preserve book identity, and fail visibly when consensus is not trustworthy.
A six-point range is a data investigation
Provenance tier: live_pregame; source table: game_odds. Jets at Steelers ranged from Pittsburgh -3.5 to +2.5 across 71 reported books, a 6-point span. Those endpoints cross zero and reverse the favored side.
That row cannot be treated as one market price. It may contain stale quotes, different capture times, incompatible home-team orientation, or genuine disagreement. The aggregate does not say which. Until the underlying book rows are inspected, any EV result built on a midpoint would be false precision.
Provenance tier: live_pregame; source table: game_odds. Panthers at Jaguars spans 4 points from -2.5 to +1.5, and Commanders at Lions spans 3 points from -4.5 to -1.5, each across 71 reported books. Those rows deserve the same quarantine-and-inspect treatment.
Smaller ranges still change the ticket
Provenance tier: live_pregame; source table: game_odds. Packers at Broncos, Saints at Rams, and Seahawks at Titans each span 1.5 points across 71 reported books.
Provenance tier: live_pregame; source table: game_odds. Giants at Dolphins, Bears at Bengals, and Eagles at Patriots each span 1 point; Ravens at Vikings spans half a point.
Those gaps may look tame beside six points, but they still describe different tickets. An EV surface should never hide them behind a consensus average. The user needs the exact spread and odds available at the selected book. Range is a warning about the aggregation, not an input to substitute for the real offer.
Matching endpoints are not automatically trusted inputs
Provenance tier: live_pregame; source table: game_odds. Bills at Browns stored -3, Falcons at Colts -3.5, Chiefs at Buccaneers -5.5, and Cowboys at Cardinals -1.5, each with matching endpoints across 71 reported books.
A zero range removes one warning but does not establish fair value. The tool still needs the odds, market type, capture time, and book identity. It also needs to know whether the same normalized observation was counted more than once. Agreement inside an aggregate is not a probability.
Provenance tier: live_pregame; source table: game_odds. The MLB rows are thinner: Braves at Brewers reports 9 books; Giants at Red Sox, Rays at Orioles, Mets at White Sox, Tigers at Royals, Guardians at Rockies, Reds at Diamondbacks, Cubs at Mariners, and Pirates at Dodgers report 8; Nationals at Marlins and Athletics at Astros report 6; Angels at Rangers reports 4. Each cited row stored matching spread endpoints.
The book count belongs on the screen because it describes market depth in the captured row. It should not be turned into a universal minimum without a tested reason. A hard threshold is product logic, and this article does not verify one.
The honest EV contract
An EV calculation needs a model probability and the exact offered odds for an exact market. A spread range supplies neither. The surface should therefore show the selected book, line, price, capture time, and any aggregation warning before it shows an edge.
When a range crosses zero or reverses the favorite, the cure is to fail closed and inspect the inputs. When a smaller range remains, let the user choose the actual offer rather than silently averaging. When endpoints match, keep showing depth and freshness instead of declaring consensus trustworthy by fiat.
The closing-line value guide explains how to judge the eventual entry, and the odds guide keeps line and price in the same calculation. This snapshot is valuable because it exposes bad assumptions. It is not evidence that any named tool already cures them.
Expected value from graded outcomes
Expected-value cells render only when a verified source binds observed win outcomes to the price paid for the same bets.
Model calibration from graded predictions
Calibration points render only when a verified source binds prediction probabilities to settled outcomes for the same observations.
