Reading an Entity Card Without Fooling Yourself

The atlas renders one small analytics card per entity across every sport in the system — 1,549 cards in total, 1,523 of them per-entity, each one traceable to a count you can reopen. Every card wears a red DESCRIPTIVE_ONLY badge, a source-floor-as-of footer, and edge_claimed=false. There is no forecast on these cards and no win probability. They show what an entity has done, gated by a declared minimum-sample floor, and nothing more. Reading one well is mostly a matter of respecting three things the card tells you about itself.

First, the floor. A card only exists if the entity cleared a stated inclusion threshold: an NBA player needs at least 800 career minutes (which admits 482 of 807 players), an MLB batter at least 300 pitches faced in 2025 (485 of 671), a tennis player at least 30 matches per surface, a soccer club at least 10 prior matches per metric. Below the floor the card shows n/a; it never fabricates a number to fill the gap. The floor is the first thing to read, because it tells you how much sample is under the pretty panel.

Second, what the rate is not. NBA player numbers are per-36-minute rates over a multi-season sample, not per-game and not opponent-adjusted. Jokic's 28.4 / 12.8 / 10.1 points, rebounds, and assists per 36 is a rate line, not a per-game average, and a per-36 triple-double is not a per-game one. The classic trap is a raw shooting percentage: Giannis shoots 60.9 percent from the field, which sounds like elite touch, but the range numbers give it away — 27.4 percent from three, 64.0 percent at the line. The high field-goal mark is mostly about where the shots come from, a rim-heavy diet, which the raw percentage cannot separate from skill.

Third, and most dangerous, the confound. Two card families look like they measure individual value and do not. The team card's top contributor is simply the player with the most total minutes, not the best player: Boston's is Derrick White at 7,576.2 minutes and Golden State's is Brandin Podziemski at 6,018.4, not Curry. And the star-removal exhibit maps a team's on-off swing onto a win-probability curve, which reads dramatically — Denver goes from 0.634 with Jokic to 0.0517 without, a 0.5822 swing — but the card labels it a ceiling, not a clean player value, because it attributes a full-lineup on-versus-off gap to one man with no control for teammates, opponent, coach trust, or garbage time. The guardrail is in the same data: the median qualified player's on-off is just +0.127, so most such gaps are noise around zero, and the biggest on-off name on Boston is the role player Luke Kornet at 0.2785, not Tatum. A big on-off number is a flag to investigate, never a verdict.

The cards are served through the same fail-closed answer engine as everything else, so asking about an entity that was never built returns an honest no_data — refusing, not guessing — rather than a plausible invention. That is the whole design philosophy in miniature: wide coverage where every unit of it is individually auditable, down to the floor that let it in, and an explicit refusal the moment you reach past what the data supports. Read the floor, respect the rate, distrust the confound, and the card will not fool you.

Sources