Interpret
Resolve a represented player, statistic, and supported scope.
Trust the answer.
Inspect the limits.
How coverage, calculations, missing information, statistical terms, and reusable interface patterns work across the product.
02 · METHODOLOGY
How answers are calculated
Natural language selects a validated query. Entity resolution, filtering, and arithmetic remain deterministic.
Validate
Reject unsupported dimensions and ask for one missing detail at a time.
Filter
Select only records matching every explicit condition.
Calculate
Compute totals, averages, highs, and percentages from qualifying records.
Verify
Keep every supporting game available from the answer.
Missing is not zero
A missing value, DNP, or unknown appearance is not inserted into an average as zero.
Shooting rates
Percentages require valid makes and attempts. Zero attempts produce an unavailable percentage.
Comparisons
Recent and previous windows are non-overlapping. Statistical differences do not imply a cause.
Score not verified
A final’s score is shown only when its box score passes three checks: the score is not a tie; each team’s minutes add up to a full game, regulation plus any overtimes, within a minute; and, in seasons that record plus-minus, each team’s players’ plus-minus adds up to five times its margin. A score that fails one reads “Score not verified”: it is not shown, and no record, series or count uses it. Finals the leagues’ own game logs and schedules show the source holds wrongly, some of which pass every check, are listed one by one and read the same way, their player lines left out, until the source is repaired and a corrected box score passes the checks.
02 · SURPRISE SCORE
How rare was that game?
The Surprise score says how rare the gap is between a game’s Game Score and the player’s own usual line, among comparable games. A 97 means no more than 3 in 100 comparable games finished that far on the same side of their usual. It says what changed, never why.
Measure
Sum the box score into one Game Score: points, rebounds, assists, steals and blocks count for the player; shot attempts, turnovers and fouls count against.
Compare
Set it against the player’s usual line, the average of this season’s earlier regular-season games. The gap is the Game Score minus the usual.
Size
Size the gap by how much the player’s own games usually vary, so a swing that is routine for one player can be rare for another.
Place
Place it among comparable games: the calibration games whose player had the nearest usual Game Score (tables below).
Label
The score is its percentile among them, rounded down: 95–97 Unusual, 98 Rare, 99 Very rare. Under 95, the game is within the usual range and shows no score.
Rarity, not size
The score measures rarity, not size. A quiet night from a player whose games rarely vary can score higher than a big night from a player whose games often swing far from their usual.
Tiers
| SCORE | TIER | COMPARABLE GAMES THIS FAR OUT | NBA · HOW OFTEN | WNBA · HOW OFTEN |
|---|---|---|---|---|
| 95–97 | Unusual | No more than 5 in 100 | About 2 to 4 times a season | About once or twice a season |
| 98 | Rare | No more than 2 in 100 | About once or twice a season | About once a season or less |
| 99 | Very rare | No more than 1 in 100 | Less than once a season | Not shownTops out at Rare |
How often: for a player who plays every game, a game this far above the usual comes about that often, and so does one this far below. Under 95, a game is within the usual range and no score is shown.
Top tier
A regular-season table reaches Very rare only when at least 50 comparable games stand beyond the 1% cut and its held-out shares at 99 or more land in 0.7–1.4% on both sides. Otherwise Rare is its top tier, and its scores stop at 98.
NBA: Very rare in both tables.
WNBA: Rare in both tables, so its scores stop at 98: they hold 3,534 and 2,692 comparable games, leaving only 35 and 26 beyond the 1% cut.
Early season
Until a player has 20 games this season, the usual line also counts last season’s average, as 10 extra games before the first game and half a game fewer after each game played, so none from game 21. From then on, the usual line is this season’s average alone. Last season counts only with 20 or more games, and not after a move to a new team or a change of 33% or more in minutes a game; without it, scoring starts after 10 games. Early-season scores have their own comparable-games table, and in a player’s first 10 games, after one game far above or below the usual, later ones on that side show no score, because the usual line may be out of date.
Which games
Scores cover regular-season games: NBA from 2023-24 and WNBA from 2022. Playoff games are scored apart: NBA from 2023-24 and WNBA from 2022. Preseason, play-in, Cup final and All-Star games are not scored. A game is scored only when the player played and every Game Score input is recorded, with points that add up from the field goals, threes and free throws; a DNP is never scored.
Zero-filled inputs
The Sport Data Service fills missing minutes, fouls, usage and true shooting with zero, so a zero there is unknown when participation is unknown. A played game with zero minutes is not scored, and a usage of zero, or a true shooting of zero with shot attempts, is read as missing. Fouls cannot be validated: a filled-in zero would count as no fouls.
Calibration
Each league’s regular-season tables are fit on earlier seasons and checked on a season the fit never saw; published rates come only from that held-out season. This fit was made Oct 19, 2026 from recorded Sport Data Service reads (9 schedules, 4,992 box scores and 3,669 game logs; none missing or unusable). It is refit each off-season.
| LEAGUE | TABLE | FIT ON | FIT SAMPLE | COMPARABLE GAMES | BEYOND 95 · 98 · 99 | HELD OUT | TOP TIER |
|---|---|---|---|---|---|---|---|
| NBA | Game 21 on | 2023-24 and 2024-25 | 32,336 games884 player-seasons | 5,000 | 250 · 100 · 50At least 50 beyond each | 2025-2616,378 games | Very rareScores reach 99 |
| NBA | Early seasonBefore game 21 | 2023-24 and 2024-25with 2022-23 and 2023-24 before them | 13,748 games999 player-seasons | 5,000 | 250 · 100 · 50At least 50 beyond each | 2025-266,721 games, 26 withheld | Very rareScores reach 99 |
| WNBA | Game 21 on | 2022–2025 | 7,068 games482 player-seasons | 3,534 | 176 · 70 · 35Under 50 beyond 99 | 20262,686 games | RareScores stop at 98 |
| WNBA | Early seasonBefore game 21 | 2023–2025with 2022–2024 before them | 5,384 games441 player-seasons | 2,692 | 134 · 53 · 26Under 50 beyond 99 | 20262,199 games, 5 withheld | RareScores stop at 98 |
Held-out check
On the held-out season, the share of scored games at or above each cut must land in its band, on each side. Held-out games are scored from the game logs, as the player pages score them; games the early-season rule leaves without a score are in no share.
| LEAGUE | TABLE | SIDE | 95 OR MORE · 4–6% | 98 OR MORE · 1.5–2.5% | 99 OR MORE · 0.7–1.4% |
|---|---|---|---|---|---|
| NBA | Game 21 on2025-26 · 16,378 games | ▲ Above usual | 4.66%763 games · pass | 1.81%296 games · pass | 0.98%160 games · pass |
| NBA | Game 21 on2025-26 · 16,378 games | ▼ Below usual | 5.18%849 games · pass | 2.04%334 games · pass | 1.07%176 games · pass |
| NBA | Early season2025-26 · 6,721 games | ▲ Above usual | 5.21%350 games · pass | 1.99%134 games · pass | 0.86%58 games · pass |
| NBA | Early season2025-26 · 6,721 games | ▼ Below usual | 4.72%317 games · pass | 2.04%137 games · pass | 1.10%74 games · pass |
| WNBA | Game 21 on2026 · 2,686 games | ▲ Above usual | 4.54%122 games · pass | 1.71%46 games · pass | 0.89%24 games · pass |
| WNBA | Game 21 on2026 · 2,686 games | ▼ Below usual | 4.43%119 games · pass | 2.23%60 games · pass | 1.23%33 games · pass |
| WNBA | Early season2026 · 2,199 games | ▲ Above usual | 4.59%101 games · pass | 1.91%42 games · pass | 1.23%27 games · pass |
| WNBA | Early season2026 · 2,199 games | ▼ Below usual | 4.59%101 games · pass | 1.64%36 games · pass | 0.96%21 games · pass |
All 24 shares land in their bands.
Playoff games
A playoff game is scored against the same season’s regular-season usual line, as its own labelled scope that is never averaged with the regular season. It is scored only for a player with 20 or more regular-season games that season, and last season never counts, so the usual line is that season’s regular-season average alone. The gap is sized as in the regular season, then placed among comparable playoff games: of the fit’s playoff games, the half whose player had the nearest usual Game Score, but no fewer than 1,000 and no more than 5,000. The tiers keep their meaning, but not the how-often column above: playoff games per player vary by team.
Playoff scores go no higher than Rare (98). A playoff table reaches a tier only when at least 50 comparable games stand beyond its cut and its held-out shares at that cut and every lower one land in their bands on both sides.
NBA: Rare, so its playoff scores stop at 98: Rare is the cap, and its 2023-24 to 2025-26 postseasons give only 26 comparable games beyond the 1% cut, where Very rare needs 50.
WNBA: Unusual, so its playoff scores stop at 97: its 2022–2026 postseasons give only 20 comparable games beyond the 2% cut, where Rare needs 50.
Why playoff tables
Scored with the regular-season tables instead, the same playoff games land in 3 of 12 bands. At 95 or more, where the design puts 5% on each side: NBA, 2.42% above usual and 8.64% below; WNBA, 3.37% above usual and 5.58% below. So playoff games get their own tables.
| LEAGUE | POSTSEASONS | SIDE | 95 OR MORE · 4–6% | 98 OR MORE · 1.5–2.5% | 99 OR MORE · 0.7–1.4% |
|---|---|---|---|---|---|
| NBA | 2023-24 to 2025-265,255 games | ▲ Above usual | 2.42%127 games · fail | 0.89%47 games · fail | 0.48%25 games · fail |
| NBA | 2023-24 to 2025-265,255 games | ▼ Below usual | 8.64%454 games · fail | 3.41%179 games · fail | 1.35%71 games · pass |
| WNBA | 2022–20261,809 games | ▲ Above usual | 3.37%61 games · fail | 1.22%22 games · fail | 0.66%12 games · fail |
| WNBA | 2022–20261,809 games | ▼ Below usual | 5.58%101 games · pass | 2.49%45 games · pass | 1.66%30 games · fail |
Playoff calibration
Each league’s playoff table is fit on the playoff games of the postseasons below, and every published share comes from held-out games. The playoff tables were fit Oct 20, 2026 from the same recorded reads of Oct 19, 2026, so a postseason still under way then counts only the games played by that day.
| LEAGUE | FIT ON | FIT SAMPLE | COMPARABLE GAMES | BEYOND 95 · 98 · 99 | TOP TIER |
|---|---|---|---|---|---|
| NBA | 2023-24 to 2025-263 postseasons | 5,255 games628 player-seasons of 20+ regular-season games | 2,627Half the fit sample | 131 · 52 · 26Under 50 beyond 99 | RareScores stop at 98 |
| WNBA | 2022–20265 postseasons | 1,809 games383 player-seasons of 20+ regular-season games | 1,000The floor: half the fit sample is 904 | 50 · 20 · 10Under 50 beyond 98 and 99 | UnusualScores stop at 97 |
Playoff held-out check
Each postseason is left out in turn and scored with a table fit on the others, and the shares are pooled, so every share comes from games its table never saw; one postseason alone is too few games to check. The table in use is fit on every postseason. The rows below also split the same games at a usual Game Score of 8.
| LEAGUE | GAMES | SIDE | 95 OR MORE · 4–6% | 98 OR MORE · 1.5–2.5% | 99 OR MORE · 0.7–1.4% |
|---|---|---|---|---|---|
| NBA | All games5,255 games | ▲ Above usual | 4.81%253 games · pass | 1.85%97 games · pass | 0.99%52 games · pass |
| NBA | All games5,255 games | ▼ Below usual | 4.68%246 games · pass | 2.00%105 games · pass | 1.01%53 games · pass |
| NBA | Usual under 82,252 games | ▲ Above usual | 4.66%105 games · pass | 1.82%41 games · pass | 1.29%29 games · pass |
| NBA | Usual under 82,252 games | ▼ Below usual | 3.20%72 games · fail | 0.80%18 games · fail | 0.44%10 games · fail |
| NBA | Usual 8 or more3,003 games | ▲ Above usual | 4.93%148 games · pass | 1.86%56 games · pass | 0.77%23 games · pass |
| NBA | Usual 8 or more3,003 games | ▼ Below usual | 5.79%174 games · pass | 2.90%87 games · fail | 1.43%43 games · fail |
| WNBA | All games1,809 games | ▲ Above usual | 5.09%92 games · pass | 2.10%38 games · pass | 1.00%18 games · pass |
| WNBA | All games1,809 games | ▼ Below usual | 4.92%89 games · pass | 2.05%37 games · pass | 1.00%18 games · pass |
| WNBA | Usual under 81,003 games | ▲ Above usual | 4.79%48 games · pass | 1.99%20 games · pass | 0.80%8 games · pass |
| WNBA | Usual under 81,003 games | ▼ Below usual | 3.49%35 games · fail | 1.00%10 games · fail | 0.30%3 games · fail |
| WNBA | Usual 8 or more806 games | ▲ Above usual | 5.46%44 games · pass | 2.23%18 games · pass | 1.24%10 games · pass |
| WNBA | Usual 8 or more806 games | ▼ Below usual | 6.70%54 games · fail | 3.35%27 games · fail | 1.86%15 games · fail |
All 12 shares of the check land in their bands. Split by usual level, 13 of 24 land in theirs; the check is on all games together.
For statisticians
z = (GS − μ) / √(s*² · (1 + 1/(n + w)))GS is the game’s Game Score; μ the usual Game Score over this season’s n earlier games plus w games of last season’s average (the early-season weight). For a playoff game, n counts that season’s regular-season games and w is 0.
s*² = (k₀·σ₀²(μ) + w·s²_last + Σ(xᵢ − m)²) / (k₀ + w + n − 1)The player’s own variance, shrunk toward the league prior σ₀²(μ) = max(a + b·max(μ, 0), 4) with k₀ = 8: xᵢ are this season’s earlier Game Scores, m their mean, and s²_last last season’s variance.
| LEAGUE | PRIOR a | PRIOR b | POOLED SD | BEFORE ROUNDING | FIT ON |
|---|---|---|---|---|---|
| NBA | 9.2 | 2.85 | 5.5 | 5.64 | 955 player-seasonsOf 15+ games |
| WNBA | 7.1 | 2.92 | 5.0 | 5.09 | 536 player-seasonsOf 15+ games |
The pooled SD is the within-player SD of Game Score pooled over those player-seasons, rounded to the nearest half point. A Game Score change smaller than one pooled SD never qualifies a page for search indexing.
The score is the comparable-games percentile of z. At each point of a grid of usual Game Scores (step 0.5), a table stores 13 quantiles of z, from 0.5% to 99.5%, among the K calibration games with the nearest μ. A score interpolates them linearly in μ, places z between them to get CDF(z), and takes ⌊100·CDF⌋ above the usual (CDF ≥ 0.5) or ⌊100·(1 − CDF)⌋ below; a table capped at Rare stops at 98, and one capped at Unusual at 97. A playoff game reads its league’s playoff table, whose K is min(5,000, N, max(⌊N/2⌋, 1,000)) for N playoff games in the fit.