Every prediction is committed with a timestamp before the match and graded in public — hit or miss. Brier score = (probability − outcome)²; 0 is clairvoyance, 0.25 is a coin flip, lower is better. The model being publicly wrong is part of the product.
committed 2026-07-12 · graded 2026-07-13 · v2 (posterior simulation), DreamBreaker treated as 50/50 · frozen source ↗
St. Louis Shock won 3-0. MXD2 skipped as a dead game; no DreamBreaker. Anna Leigh Waters lost both her games (6-11 WD, 8-11 MXD1).
| call | prob | outcome | grade | Brier |
|---|---|---|---|---|
| Overall: St. Louis Shock wins the final | 61% | STL won 3-0 | ✓ HIT | 0.154 |
| WD: Waters/Johnson over Bright/Fahey | 88% | Bright/Fahey won 11-6 | ✗ MISS | 0.774 |
| MD: Tardio/Patriquin over Khlif/Howells | 92% | Tardio/Patriquin won 11-3 | ✓ HIT | 0.006 |
| MXD1: Waters/Khlif over Bright/Patriquin | 54% | Bright/Patriquin won 11-8 | ✗ MISS | 0.292 |
| MXD2: Fahey/Tardio over Johnson/Howells | 75% | not played (match decided 3-0) | – VOID | — |
| Match reaches a DreamBreaker | 46% | no DreamBreaker | ✓ HIT | 0.216 |
The earlier v1 forecast (same file, committed hours before) grades identically in direction: overall STL 57.0% HIT, WD 69.5% MISS, MD 79.1% HIT, MXD1 52.9% MISS. Both models called the match and missed the headline WD game.
committed 2026-07-12 · graded 2026-09-02 · to be graded after 2026-09-01 · v1 dyad chemistry + fixed-effects regression scoring (method frozen in the source file) · frozen source ↗
Scored on 2,326 modelling games dated after 2026-07-12 (full 2026-season player fixed effects, H2-only dyad estimate, cluster-robust by match). Both registered PAIR predictions are VOID, not wrong: Waters/Bright and Alshon/Patriquin never played together again after the freeze. All four players were active in H2 (25-33 games each), so this is dissolution, not missing data - and it is the outcome the survivorship work predicts for the two most negative chemistry contrasts on record. The two scorable predictions both HIT.
| call | prob | outcome | grade | Brier |
|---|---|---|---|---|
| Waters + Bright (womens, PPA) underperform sum-of-parts in remaining 2026 games | — | 0 games together after 2026-07-12 (Waters 25 H2 games, Bright 26 - but never as a pair) | – VOID | — |
| Alshon + Patriquin (mens, PPA) underperform sum-of-parts in remaining 2026 games | — | 0 games together after 2026-07-12 (Alshon 33 H2 games, Patriquin 26 - but never as a pair) | – VOID | — |
| Bright + Patriquin (mixed) chemistry stays small: |chem| < 1 point/game unshrunk | — | +0.05 +/- 2.17 pts/game on 14 H2 games together (full-season +0.29 +/- 0.85); |chem| well under 1 | ✓ HIT | — |
| League-wide dyad-chemistry scale (sd_d) refit on H2-only games stays under 1.0 point/game | — | sd_d = 0.340 +/- 0.248 refit on the 2,326 H2 games (~95% upper 0.835); frozen full-season value 0.473 | ✓ HIT | — |
Scored 2026-09-02 against games dated after 2026-07-12 only, using the method written into the frozen file. Two of four predictions were VOIDED by roster churn: a chemistry prediction about a pair needs the pair to keep playing. Future preregistrations should name a fallback or a minimum-games condition.
committed 2026-07-14 · graded 2026-07-16 · v2 current values, projected lineups, singles-model DreamBreakers, calibrated · frozen source ↗
Favorites went 4/5. Columbus swept both its matchups 4-0; SoCal and Phoenix won as modest favorites; the miss was Phoenix-Orlando, where 61% Phoenix lost 2-3 on a DreamBreaker the model had rated at 37%. Jack Sock (in for the projected Yates Johnson) played both Orlando matchups.
| call | prob | outcome | grade | Brier |
|---|---|---|---|---|
| Matchup 1: Columbus Sliders over Orlando Squeeze | 99% | Columbus won 4-0 | ✓ HIT | 0.000 |
| Matchup 2: SoCal Hard Eights over Florida Smash | 63% | SoCal won 4-0 | ✓ HIT | 0.134 |
| Matchup 3: Phoenix Flames over Orlando Squeeze | 61% | Orlando won 3-2 (DreamBreaker 21-14) | ✗ MISS | 0.377 |
| Matchup 4: Columbus Sliders over California Black Bears | 99% | Columbus won 4-0 | ✓ HIT | 0.000 |
| Matchup 5: Phoenix Flames over Florida Smash | 63% | Phoenix won 3-1 | ✓ HIT | 0.138 |
Matchup mean Brier 0.130 (coin flip = 0.25). Game level, matched on the forecast's own team-one pairing: 12/13 hits, mean Brier 0.092, two modal scores exact (both 11-6 mixed calls); 7 games VOID where the projected pairing never took the court, and opponent-side substitutions (Sock for Johnson, twice) graded through as lineup risk. Reaches-a-DreamBreaker graded 4/5 on the modal call, mean Brier 0.136 — the miss was the 37.4% DB that decided Phoenix-Orlando. POST-HOC COUNTERFACTUAL (not a regrade; committed grades stand): re-rating the four Sock-affected games with actual lineups (Sock 0.837 vs Johnson 0.369 on 448/303 games) flips the matchup-3 call to Orlando 68% (Brier 0.377 → 0.102) and dissolves the MD game miss (0.316 → 0.023) — the day's headline miss was lineup information, not skill values. It cuts the other way too: Columbus's wins over the Sock-rated pairs (11-6, 11-1) become 48% and 90% calls, so the affected games' Brier sum rises 0.362 → 0.507. The model can't take credit both ways; the ledger keeps the frozen numbers.
committed 2026-07-17 · graded 2026-07-17 · OLS with cluster-robust SEs: next-10-rally point differential, receiving-side timeout vs untreated moments, adjusted for pre-window differential, margin, server number, strength · frozen source ↗
No meaningful windowed timeout effect: theta = +0.060 points per 10-rally window (99% CI [-0.034, +0.155], p=0.10) - BOUNDED under the 0.5-point threshold. Secondary 5-rally window +0.044 [-0.017, +0.105]. The trigger profile confirms the intuition about WHEN timeouts happen (called after an average -2.6-point run vs -0.25 at ordinary moments), and the naive after-minus-before comparison shows +2.30 points of apparent improvement - almost exactly the +2.06 a zero-effect simulation with endogenous calling produces. The perceived timeout turnaround is regression to the mean.
| call | prob | outcome | grade | Brier |
|---|---|---|---|---|
| H3b: no meaningful windowed timeout effect (|theta| < 0.5 pts per 10 rallies or p >= .01) | 75% | +0.060 pts/10 rallies, 99% CI [-0.034, +0.155] - no meaningful effect, BOUNDED | ✓ HIT | 0.062 |
User-motivated refinement of H3 (strategic adjustments may express over several rallies, not one). The naive before-vs-after comparison is demonstrated to be biased under the null by the endogenous-timeout selftest; the registered comparison is treated vs untreated moments of the same kind.
committed 2026-07-17 · graded 2026-07-17 · Same matched-moment design as H3b; outcome = receiver's share of next 10 rallies won. Secondaries: 3-rally share, possession-killer, severe-trigger subset. Any positive requires disjoint-sample replication before any claim. · frozen source ↗
Bounded null at the steelman's strongest: receiving side wins +0.20pp more of the next 10 rallies after its timeout (99% CI [-0.30, +0.71], p=.31; threshold was 2.5pp). Possession-killer dead null (-0.05pp); severe-trigger subset -0.11pp; 3-rally share +0.86pp at p=.045 (lone sub-threshold secondary; noise per the registered language). The naive before/after illusion SCALES with desperation - +16.4pp overall, +25.5pp for severe triggers - while the adjusted effect stays at zero: the more the timeout was needed, the bigger the regression-to-the-mean mirage. Timeout program closed on this sample; replication pool reserved but nothing to replicate.
| call | prob | outcome | grade | Brier |
|---|---|---|---|---|
| H3c: no meaningful rally-share timeout effect (|theta| < 2.5pp or p >= .01) | 75% | +0.20pp of next-10-rally share, 99% CI [-0.30, +0.71] - no meaningful effect, BOUNDED | ✓ HIT | 0.062 |
User-proposed steelman: rallies are the proximal outcome (killing the server's possession precedes scoring, which point differentials cannot see until converted). Declared the FINAL timeout estimand on this sample; replication pool (untouched ~10k eligible matches, seed 20260718) reserved for any positive.
committed 2026-07-17 · graded 2026-07-17 · Window profile theta_W for W in {1,2,3,4,5,7,10} on the full untouched 9,426-match pool, identical moments/controls. Primary: theta_3 > 0 one-sided alpha .01. Family: other windows at Bonferroni |z|>=2.69. Bounding: all seven 99% CIs vs +/-2.5pp. Amended pre-unblinding (user's argument: one cell cannot close a family for a skeptical reader in either direction). · frozen source ↗
The 3-rally hint was noise. On the full untouched pool (9,426 matches, 361,593 moments — 4x the original), the primary window-3 effect is +0.26pp (one-sided p=.113, fails to replicate at .01), no other window clears Bonferroni (max |z|=1.27), and all seven 99% CIs from window 1 to window 10 sit inside +/-2.5pp. The window profile is flat and incoherent, not the decay a real short-window huddle effect would show. Frozen conclusion: no window from 1 to 10 rallies hides a meaningful timeout effect. Timeout question closed in both directions.
| call | prob | outcome | grade | Brier |
|---|---|---|---|---|
| H3d primary: the 3-rally hint FAILS to replicate (one-sided p >= .01) | 80% | +0.26pp, one-sided p=.113 — fails to replicate (predicted) | ✓ HIT | 0.040 |
| H3d family: no window in {1,2,4,5,7,10} clears Bonferroni |z|>=2.69 | 80% | no window clears Bonferroni |z|>=2.69 (max 1.27); family bounded inside +/-2.5pp | ✓ HIT | 0.040 |
Hypothesis generated on sample A (lone p=.045 secondary among ~10 estimands), tested once on disjoint data. Even a successful replication claims only a real-but-tiny sub-1pp huddle effect, far below the 2.5pp meaningful bar.
committed 2026-07-17 · graded 2026-07-17 · cluster-robust logistic regressions on 191,500 resolved rallies (2,500 matches, 0 dropped); serve state + v2 strength controls; alpha .01 + practical thresholds · frozen source ↗
Momentum is not detectable in pro pickleball rallies. Within-possession streaks +0.11pp (99% CI [-0.72, +0.94], bounded null), trailing-5-rally momentum +0.17pp ([-0.19, +0.52], bounded null), game-point pressure -0.01pp ([-1.31, +1.29], bounded null). Timeouts: inconclusive, direction mildly AGAINST the caller (+1.53pp for the opponent's next serve rally, p=.014, CI reaches +3.1pp).
| call | prob | outcome | grade | Brier |
|---|---|---|---|---|
| H1: no meaningful within-possession momentum (|APE| < 1pp or p >= .01) | 85% | +0.11pp, 99% CI [-0.72, +0.94] — no meaningful effect, BOUNDED | ✓ HIT | 0.022 |
| H2: no meaningful cross-possession momentum (trailing-5 window) | 85% | +0.17pp, 99% CI [-0.19, +0.52] — no meaningful effect, BOUNDED | ✓ HIT | 0.022 |
| H3: no meaningful receiving-side timeout effect (|APE| < 2pp) | 80% | +1.53pp, p=.014 (not sig at .01); CI [-0.07, +3.12] exceeds +2pp so not bounded — no effect established, formally inconclusive | ✓ HIT | 0.040 |
| H4: no meaningful game-point pressure effect (|APE| < 2pp) | 70% | -0.01pp, 99% CI [-1.31, +1.29] — servers convert game points at exactly their usual rate, BOUNDED | ✓ HIT | 0.090 |
4/4 registered predictions hit, mean Brier 0.044. Secondaries: lag-1 rally -0.18pp (ns); serving-side timeouts -3.24pp on only 512 events (ns, CI +/-5.5pp). Descriptive bonus: timeouts are called by the RECEIVING side 12.5x as often as the serving side (6,417 vs 512) - they are momentum-stopping attempts by the team being scored on, and the next-rally evidence says they do not detectably work. Caveats as registered: logged matches are not a random sample of events, and the timeout estimate admits a selection interpretation; under the iid null the tests remain exact.
When the frozen model said a team should win X% of the time, how often did they actually win? Dots on the diagonal = honest probabilities. Curve measured out-of-sample: every post-June-2026 game rated by the frozen pre-June fit (4494 games). The fitted correction is nearly the identity (slope 0.95) — the earlier v1 model ran underconfident, this one doesn't. What the data does insist on: across all 36k games, favorites rated above 99% still lost 52 of 5016 games (~1%), so site probabilities are floored — nothing is ever 0% or 100%. Frozen predictions in the ledger above are graded exactly as committed, never re-calibrated after the fact.