Experimental measurements
Each measurement was searched against prior art before it was built, re-derived to try to break it, and published with the confounds that could change its interpretation.
1,307,128 move pairs
soccer_intl completes half its pre-game line motion by 2.7h before tip; nba not until 4.8h.
Limitationbucket_move sums ABSOLUTE per-tick moves -> this is churn/activity, not net price travel; back-and-forth noise and correlated consensus books inflate it.
174 games
mlb stays contested later (LCF 0.737) than soccer_intl (LCF 0.599).
Limitationthresholds are NOT truly equated across sports (2 runs != any goal margin != any NBA margin) and clock units differ (innings vs minutes), so the cross-sport ranking is indicative, not metric.
30 teams
Most one-star-fragile: DEN (Nikola Jokić, delta_winprob 0.5822).
Limitationboth estimators carry the ROSTER CONFOUND both source files label explicitly: why a player missed is entangled with injury clustering, rest, blowout pulls, and replacement lineup -- none controlled.
19 in-game checkpoints
mlb: MFP rises from 0.05 to 0.41 over the game (mean 0.15); soccer_intl: MFP rises from 0.68 to 1.49 over the game (mean 0.92)
Limitation'model' is one specific state-only forecaster, so MFP conflates genuine news the closing reference forecast prices (lineups, momentum) with plain model misspecification.
4,406 max-disagreement observations
soccer_intl: overshoot 0.286 but model matches or improves on the reference only 22% of the time at max disagreement -> OHG 0.081 (a failure to harvest, not an advantage); mlb: overshoot 0.072 but model matches or improves on the reference only 38% of the time at max disagreement -> OHG 0.009 (a failure to harvest, not an advantage)
LimitationMANDATORY: OHG is an honest NULL -- at max disagreement the closing reference forecast is usually right (model_closer_rate 0.38 mlb / 0.22 soccer).
511,807 adjacent pitch pairs
95 percent interval [0.0348, 0.0385]
Published preregistered claims: 1 confirmed; 2 contradicted. Confirmed claim: The repeat excess is positive overall. Contradicted claims: The repeat excess is larger in pitcher-ahead counts than in counts where the pitcher is behind; Repeated pitches draw a lower rate of swings and misses than switched pitches at the same count.
LimitationDescriptive conditional frequencies, not a forecast and not a causal estimate.
4,793 NBA games
95 percent interval [0.0257, 0.1502]
The differential tracks the frequency; the two symmetric-congestion groups look alike, which is no visible difference rather than an absence of fatigue.
LimitationDescriptive conditional frequencies, not a forecast and not a causal estimate.
90 team-seasons with published schedule counts
The claimed ORtg tax uses a scoring-margin contrast as its anchor. The ORtg effect requires recomputation before it can be reported.
LimitationHistorical source artifact is available for inspection. Its separate per-36 composite remains raw descriptive data.
52,078 MLB team-starts
95 percent interval [0.4984, 0.5069]; [0.497, 0.5077]; [0.4914, 0.5079]
Recorded as a null.
LimitationDescriptive conditional frequencies, not a forecast and not a causal estimate.
Publication checks
The definition is searched against published literature and the public analytics canon before it is built. The verdict remains visible when someone reached most of the way there first.
The stat must be computable from committed artifacts. No new scrape, private feed, or number that cannot be recomputed from its cited source files.
The result is rebuilt by a different route where one exists, specifically to try to break it. Disagreement between methods is printed rather than hidden.
Every stat publishes what could make it wrong: sparse tail buckets, excluded sports, absolute-move churn, or model misspecification. Confounds are part of the stat.
These are measurements of observed structure and model behavior. Each artifact states its result and limits together.
These measurements are explicit about their own limits, and every one can be recomputed from the artifacts named on its card.