PRESEASON ANALYSIS
2025–26 · MEASURED PERFORMANCE
AF-Score 77.3 placed San Antonio Spurs #1 in the complete 30-team 2025–26 field. AF-Score v1.1 is rebuilt from 100% observed regular-season team boxes and remains a descriptive performance index, not a win probability.
2025–26 FULL-FIELD LEADER77.3AF-Score · 0–100 performance index100% observed team boxesFIVE DIMENSIONS
AF-Score is not one hidden number. Different teams can lead different parts of measured performance. These component values are index scores, not probabilities.
2025–26 · PERFORMANCE MOVEMENT
Boston Celtics moved from 49.0 after Game 10 to 69.6 at the end of the regular season. The largest arithmetic contribution came from Scoring +8.8 AF-Score points.
Contribution describes the index arithmetic. It does not claim that scoring caused wins or future outcomes.
See the full Movement analysis →OTHER SIDE OF THE MOVE
Milwaukee Bucks moved from 49.8 to 29.4. Its largest negative contribution was Scoring −7.7 AF-Score points.
The page compares the same complete 30-team field at both endpoints.
Compare all risers and fallers →SEVEN-SEASON POWERHOUSES
Boston Celtics leads the seven-season archive with 5 Top-5 finishes, followed by Denver Nuggets with 4. Every season now uses a complete 30-team full-observed field.
Boston led the complete AF-Score field three seasons in a row: 2022–23, 2023–24 and 2024–25. AF-Score v1.1 now supports this as a full 30-team historical statement in each season.
SEVEN-SEASON AF-SCORE ARCHIVE
Milwaukee Bucks72.530/30 full observed2020–21
Utah Jazz69.930/30 full observed2021–22Phoenix Suns72.830/30 full observed2022–23Boston Celtics69.230/30 full observed2023–24Boston Celtics78.730/30 full observed2024–25Boston Celtics76.430/30 full observed2025–26San Antonio Spurs77.330/30 full observedGAME-WIN MODEL · FORMAL OOS
Across 5,282 games held out from the final model-selection decision, ProbHoops produced 39 more correct winner calls than Classic Elo and lower Log Loss in 4/4 OOS seasons.
WHY ACCURACY ALONE IS NOT ENOUGH
Classic Elo had slightly higher raw winner accuracy that season — 65.56% vs ProbHoops 65.33% — while ProbHoops still recorded the better Log Loss: 0.6117 vs 0.6187.
That is why promotion uses proper probability scores and stability, not a single headline accuracy percentage.
See all four temporal folds →PLAYER-AWARE MODEL · SHADOW OOS
The player-aware challenger used only information available before each game. It did not consume the game’s observed minutes, retrospective DNP labels or inferred injury reasons.
Why it is not live yet: the challenger remains shadow-supported. Current probabilities stay on the accepted champion until current roster-to-role binding and the expected-minutes allocator pass their own operational gate.
PLAYER MILESTONES · SEPARATE MODEL
The semantics are exactly P(milestone reached | player is active in the game). This is not game-win accuracy, and current 2026–27 milestone probabilities remain withheld until deterministic pre-tip availability exists.
2026–27 CURRENT ANALYSIS
No current AF-Score, performance movement or player-milestone row is fabricated from last season. When verified 2026–27 evidence exists, these same surfaces can switch from preseason history to current intelligence.
PROBHOOPS BRAIN · GOVERNANCE
The historical stories above are deterministic and require no OpenAI call. The Brain remains downstream of canonical current-season facts and can narrate only what a governed public release already proves.
Evidence contract.
AI is downstream of deterministic data.
No API key or provider call is exposed to the browser.
Loading narrative contract…
CURRENT-SEASON DETERMINISTIC INTELLIGENCE
Single-pass conditioned scenarios.
Paired home-win / away-win probability swing.
NBA-official, pregame, paired counterfactual impact.
Current intelligence remains unavailable until the exact forecast and deterministic counterfactual gates are active.
Toronto Raptors98.0
Dallas Mavericks89.3