Browse the analytics

Zero of 30 teams clear the floor of at least 2 games both leading AND trailing at halftime in this 26-game corpus, so a per-team front-runner-vs-comeback profile is NOT buildable; the only robust readout is league-level — second-half margin regresses on halftime margin with a slope of 0.2885.

Of the 50 highest-minute players, the most context-sensitive shooter is Payton Pritchard, whose True Shooting % swings by only about 4.4 points on average across defense tier, venue, and rest — his biggest single split is +6.0 TS points on the second night of a back-to-back.

Across 408 qualified players, a team's win rate with the player active versus missing swings by only +0.0087 on average — the extremes are loud (Brandon Miller: Charlotte 63.1% with him, 17.6% without) but small-sample and confounded.

A time-decayed in-game prior transfers cleanly across tours: fit on WTA and tested on ATP, held-out Brier improves from 0.155 to 0.1432 (delta 0.01184), verdict REPLICATED, while a surface-specific hold prior is REJECTED.

Take Nikola Jokic off Denver and, on a neutral floor against a league-average team, the Nuggets' win probability falls from 0.634 to 0.0517 — a 0.5822 swing, the largest of any team's top on/off player.

On this international-soccer corpus the market is sharper than the model: model Brier is 0.227887 versus market 0.142726 over 9,003 rows, and the gap is widest exactly where the two disagree most.

At the widest model-market disagreement (gap >=.10, n=33402 MLB ticks) the model is the closer forecaster only 0.3773 of the time, and its Brier balloons to 0.282705 versus the market's 0.210258.

Across the NBA's observed pace range (about 92 to 104 possessions), an 80% favorite's upset probability rises only from 19.5% at the fastest pace to 20.9% at the slowest — a 1.4-point swing.

From 77,744 box-score rows the system can honestly build box-only value and unadjusted on/off, but RAPM, EPM, DARKO, BPM and RAPTOR are all marked do-not-fake because the inputs do not exist.

A frozen box-score value index puts Nikola Jokic on top at 30.86 points of value per 36 minutes, ahead of Giannis Antetokounmpo (27.558) and Victor Wembanyama (26.261).

In the NBA, roughly a third of all pregame line movement lands in the final hour before tip — final_hour_movement_share 0.3247, the highest of any measured sport.

In-game MLB total-market quotes on Kalshi update at a median of 35.0 seconds between ticks, and each move is small — a median 0.02 change in market probability.

Under a single shared reliability methodology, only NBA's market map clears the honesty floor — and there the market's in-game ECE is 0.0064 across 1593 games.

The answer engine passes 87/87 regression-bank checks and honestly answers 36.6% of answerable questions — refusals are the fail-closed design, not a defect.

For the WNBA, the devigged market's Brier score improves as the game nears — from 0.245445 six hours out (T-6h) to 0.228045 at the close, across 30 games.

Only 2 of 4 sports have a market-side reliability number, and in both the model is less reliable than the market — MLB gap 0.00664, soccer 0.039414.

628 REJECT-family rows sit in the graveyard against only 95 SHIP verdicts — a reject is honest market-efficiency evidence, not a failure.

Across 259 claim families the nulls-or-worse (134) slightly outnumber the 121 verified, an honest ratio of 1.107 nulls per confirm.

Only about half of testable mechanisms survive — overall survival rate is 0.5039 across 256 testable hypotheses.

Nulls (351) outnumber confirms (168) 2.1x — we publish our nulls.

Pooled over 74,450 player-games, the home advantage at the box-score grain is small and lives mostly in shooting: home teams shoot 47.479% from the field versus 46.833% on the road (+0.647 points) and 36.632% versus 35.867% from three (+0.765 points), while the typical player's home-minus-away scoring gap is just +0.04 points per game.

Pooled over three seasons and 870 team pairings, Atlanta versus Indiana is the highest-scoring matchup at a mean of 255.7 combined points over 10 meetings, while Charlotte versus Portland is the lowest at 197.0 — a 58-point gap between the league's fastest and slowest recurring games.

soccer_intl: overshoot 0.286 but model beats market only 22% of the time at max disagreement -> OHG 0.081 (a failure to harvest, not an edge); mlb: overshoot 0.072 but model beats market only 38% of the time at max disagreement -> OHG 0....

After an in-game price move, MLB prices land a few points above the eventual outcome (moved-to-minus-outcome +0.0518 to +0.0801 across move sizes), while in soccer the moved-to price sits well below it (-0.1945 to -0.3648).

Of 579 qualified players, Luka Doncic is the most consistent game to game, with a shrunk composite coefficient of variation of 0.3448 — well under the league median of 0.6235 — edging out Nikola Jokic at 0.3456.

In MLB, the market's Brier stays below the model's at every inning checkpoint, and the gap widens as the game progresses — market_minus_model_brier goes from -0.015 in the 1st inning to -0.0902 by the 9th.

The model's Brier Skill Score versus the market is negative across every sport — MLB overall -0.1502, soccer -0.5967 — so the model does not beat the market, and that null is the exhibit's whole point.

Ranking 153 shared national teams by trailing-10 form win-rate versus net-xG strength gives a Spearman rho of 0.6419 — moderate agreement between two independent lenses, not a temporal persistence test.

For a 1-to-5 point lead with 24-36 minutes left, the model's mean win probability is 0.6204 across 898 games and teams in that state actually won 62.03% of the time — a near-perfect calibration match.

In the 2025-26 slice, Victor Wembanyama posts the biggest on/off net-rating swing at +17.043 per 48 — yet the median qualified player sits at just +0.127, so most on/off gaps are noise around zero.

Pitchers repeat themselves most when they are behind: after a four-seamer in a three-ball count the very next pitch is another four-seamer 57.85% of the time, versus 43.23% across all counts.

Fall behind 3-0 and the pitcher almost always reaches for the heater: on a 3-0 count a four-seam fastball is the pitch 65.12% of the time, thrown in the strike zone 62.93% of the time.

Victor Wembanyama is the league's biggest form riser: his 10-game rolling per-36 box composite climbed +19.821 over the observed span, ending at 37.869 — the 100th league percentile.

On the second night of a back-to-back, NBA per-36 box-composite production dips about 0.315 versus rested games (13.591 vs 13.905) — small, but it stacks with schedule compression.

Most-taxed schedule: DEN 2025-26 (-0.38 pts/100 ORtg season-averaged); least: MEM 2025-26 (-0.27). League schedule-inequality range 0.11 pts/100 across 90 team-seasons.

The most extreme clay-over-hard specialist in the ATP career snapshot is Luciano Darderi, a 0.4603 clay-minus-hard win-rate gap, 0.6825 on clay against 0.2222 on hard.

Most one-star-fragile: DEN (Nikola Jokić, delta_winprob 0.5822). Two estimators name the same #1 player for 1/30 teams (directional cross-check, different seasons).

The full ledger exports 287 tested mechanisms — 130 confirmed, 126 null, 31 not testable — each with its effect size, p-value, and corpus pointer.

The MLB model's calibration error fell from 0.1165 in June to 0.0834 in July, yet the market stayed sharper in both months (0.0944, then 0.0635).

soccer_intl completes half its pre-game line motion by 2.7h before tip; nba not until 4.8h. tennis, wnba move most of their line >6h before tip.

The 2025 Statcast pull holds 693,037 pitches, and the framing ingredients (plate location, zone edges) are present on 99.63% of them.

The NBA drifted slightly less balanced across three seasons: the win-share Gini rose from 0.177 in 2023-24 to 0.2004 in 2025-26.

mlb: MFP rises from 0.06 to 0.24 over the game (mean 0.25); soccer_intl: MFP rises from 0.64 to 1.71 over the game (mean 0.92)

The computer-vision tracker has produced 17,254 feature rows across 241 games and 252 players, a volume readout only.

3,224 commits from 2026-03-09 to 2026-07-22, 95.75% of them agent-authored — the build itself is the product.

mlb stays contested later (LCF 0.833) than soccer_intl (LCF 0.748). (nba comeback-rate cross-check 0.0146)

By the 9th MLB checkpoint the model's forecast uncertainty has collapsed to 0.3656 bits while the market still sits at 0.7244, and their disagreement widens from 0.07 early to 0.2791 late.

In MLB the model's reliability is within 0.0066 Brier of the market's, but its resolution trails by 0.0235 — the shortfall is missing information, not miscalibration.

The median four-seam fastball in the 2025 pull sits at 94.5 mph, with 34.78% of four-seamers landing in the 95-98 mph band and 8.0% in 98-101.

1,249 player dossiers span 28 analytic categories, but the median dossier is only 46.4% complete.

We graded the bookmakers on their own devigged accuracy: Pinnacle is barely sharpest on tennis match-winner (0.1975 vs Bet365's 0.1980), but on soccer over/under the books are a statistical dead heat — no public book is meaningfully bea...

Descriptive MLB Statcast-derived leaderboards (catcher/umpire out-of-zone strike rate, platoon on-base splits) for a fixed 2022-2023 corpus slice, shown beside two honest nulls — an unstable park-factor read and a REJECTed umpire-totals...

Structural fatigue, rest, and clutch momentum effects in the NBA are confirmed — and some independently replicated — while the individual game-to-game hot/cold carryover shapes (a player's own B2B dip, foul-trouble shifts, rest-differe...

The largest single calibrated win-prob swing between adjacent buckets is a soccer collapse: a team at .4-.6 in minutes 15-30 that falls to .2-.4 by minutes 30-45 loses 0.7683 of realized win probability (min support n=120).

Greater than the sum of their parts: which five-man lineups beat what their individual on/off numbers predict — the Grizzlies' Jackson-Morant-Bane-Edey-Wells five topped 2024-25 at +24.32 net per 48 above expected.

The late-game per-36 SHIFT analytic is NOT BUILT: the NBA boxscores parquet is a full-game aggregate with no period or clutch grain, so a Q4/OT split cannot be computed from it without fabricating the numbers.

Aging curves are not buildable from this data — no birthdate or age column exists in any NBA table and only 3 season snapshots are on hand, so the honest move is to not pretend we have a curve.

Tennis momentum is real point-to-point but dies at the game grain; two tiebreak myths are null; altitude and travel effects are confirmed (altitude replicated across two disjoint year slices).

Descriptive Q4-vs-Q1-Q3 per-36 scoring, rebounding, and assist rate shift per player, from local per-quarter boxes — not a late-game-skill or predictive claim.

Home advantage is real; the true-home/neutral gap grew not because true-home rose (it is flat) but because neutral-venue advantage collapsed over time.

The betting market is well-calibrated: MLB moneyline is essentially efficient, tennis shows only a mild favorite-longshot bias.

Small-sample rate leaders regress toward the group mean under empirical-Bayes shrinkage; the raw leaderboard overstates them.

Who makes opponents stop attacking the rim: a transparent on/off rim-deterrence leaderboard — and the eye test agrees.

5 claim families changed their verdict as more data arrived; here is each mind-change, in sequence.

mlb: 78,986 residual rows carry the independent information of at most ~227 games.

composition only — each card copied from its module's own novel_*.json output

Across 178 MLB games, a 2-run gap becomes permanent by a median of inning 6.0 in 75.3% of games, while a 5-run gap only locks in by a median of inning 7.5 and in just 27.0% of games — bigger cushions clinch later and rarer.

The MLB run-total market carries 2113 rescued in-game ticks but no resolved outcomes, so its Brier is reported null by design — only the median model-vs-market divergence (0.116128) is disclosed, never a fabricated score.

Resampling whole games 1000 times, 0 of 10 MLB market probability bins deviate from the diagonal beyond noise, while the model misses in 4 of 10 — worst at the top, where 0.9541-favorites win just 0.6632 (gap -0.2909).

Across 17 head-to-head checkpoints the model out-sharpens the market in only three, all MLB late-inning run-totals (end innings 6, 7 and 8), and every one is still flagged PROVISIONAL.

83 paper bets were logged, all executed=False, and realized CLV is unmeasurable for every settled row — the ledger records 'no close captured' rather than fabricating a number.

Within an MLB game, the model's win-probability errors are almost frozen from one row to the next — the median lag-1 residual autocorrelation is 0.9809 across 222 games.

The single largest slice of the model's cumulative error is early-inning coin-flip games — MLB innings 1-3 at a .4-.6 model probability carries 6063.25 in total absolute residual mass (mean_abs_residual 0.4912 over 12,344 rows).

Late in MLB games, when the model calls a team 0.9211 to win, it actually wins only 0.6853 of the time — a 0.2357 calibration error and the model's single biggest well-sampled miss (n=5037).