DiamondWorld

Data and evaluation artifacts for DiamondWorld, a plate-appearance-level baseball world model. The headline metric is cross-player rate correlation: how well the model's simulated season reproduces the K, BB, Hit and HR rates of individual batters with at least 150 plate appearances in a held-out season.

What is here

Path	Size	What it is
data/processed/pitches_YYYY.parquet	183 MB	Pitch-level records, 2015-2024, one file per season. This is what training and evaluation read.
data/processed/validation_YYYY.json	small	Per-season row counts and schema checks from the ingest.
data/eval2/	72 MB	Evaluation outputs: paired bootstrap results, per-model player-correlation arrays, calibration and backtest reports.
data/chunks/	1 MB	Per-chunk simulation checkpoints from the pre-game sweep.
data/projections_2024.csv	159 KB	Public projection-system baseline used for comparison.
RESULTS.md	89 KB	The full experiment log, including refuted ideas and corrections.

Optional tiers, uploaded separately: checkpoints/ (630 MB of trained parameters) and data/raw/ (32 GB of API and Statcast pulls, packed as archives).

Read this before using the numbers

RESULTS.md opens with a correction banner, and it is not decorative. An external review in August 2026 found three defects that had been silently inflating results:

1. A metric that could not fail. p0_error computed P(runs >= 0), which is 1 by construction, so a shutout-rate gap always scored as zero error.
2. An unknown-player sink. Embedding index 0 was not reserved for unseen players; it was the first real player in the training table, so every unseen player inherited that player's learned representation, and index 0 was then scored as if it were a batter with ~11,800 plate appearances.
3. Bullpen leakage. The "pre-game" simulator took each team's relievers, and their appearance order, from the completed game.

Defects 1 and 2 are fixed and every affected number in RESULTS.md has been recomputed. The corrected headline is 0.624 average cross-player correlation (K .792, BB .651, Hit .445, HR .610), not the 0.611 reported earlier. Several previously reported improvements did not survive: the v21 gain fell from +0.025 (p=0.054) to +0.015 (p=0.130), and three ablation verdicts flipped.

Defect 3 affects game-level results only; the plate-appearance-level numbers above are unaffected. The leak-free sweep has now been run, and the result is worth stating plainly: with the leak removed, the simulator's win probabilities are worse than a constant home-field base rate (log-loss 0.6985 vs 0.6923), where the leaky arrays had them better (0.6881). Log5, which uses nothing but season win rates, gets 0.6709. The run-total distribution is unaffected by the leak and is where the model genuinely performs: it reproduces real overdispersion (1.98x independent-Poisson variance against a real 2.11x), though a league-wide negative binomial with no team information is better calibrated still.

A power analysis over the same test season puts the minimum detectable effect at about +0.030 average correlation at 80% power. Differences smaller than that are not resolvable with one season of held-out data, whatever their point estimate.

Attenuation ceilings

Observed rates are noisy, so correlation against them is bounded. Method-of-moments reliability gives per-rate ceilings of K 0.929, BB 0.852, Hit 0.691, HR 0.794, and 0.816 on average. The model is well short of those, so this is not a saturated benchmark.

Provenance and licensing

Derived from MLB Advanced Media game feeds and Statcast. This upload is a research artifact; the underlying data is MLBAM's and is subject to their terms. Check those terms before redistributing, especially the raw tier.

Code

https://github.com/lblommesteyn/DiamondWorld