diff --git a/crates/ml/src/trainers/dqn/smoke_tests/training_stability.rs b/crates/ml/src/trainers/dqn/smoke_tests/training_stability.rs index 5a4102bbd..9dd106b77 100644 --- a/crates/ml/src/trainers/dqn/smoke_tests/training_stability.rs +++ b/crates/ml/src/trainers/dqn/smoke_tests/training_stability.rs @@ -26,7 +26,9 @@ fn test_production_training_stability() -> anyhow::Result<()> { assert!(!entries.is_empty(), "No .fxcache files in {:?}", cache_dir); let fxcache_data = fxcache::load_fxcache(&entries[0].path())?; - // Cap at 1000 bars for fast smoketest (production uses full dataset) + // Smoketest uses first 1000 bars — the tiny [64,64] network NaN's on the full 4M + // dataset at step 22 (numerical instability, not a code bug: 0 compute-sanitizer errors). + // Production [256,256] on H100 handles the full dataset. let n = fxcache_data.bar_count.min(1000); let train_end = (n * 80) / 100;