RESULTS.md — Verified Paper Metrics Generated by analysis/produce_paper_metrics.py. All numbers in the paper should come from here.
Checkpoint: model_ckpts/model_final_30_conditioned.pth
Test set: N = 39189 samples
Extinction threshold (for generated): θ = 0.005
Metric
Value
95% CI
$R^2$ normalized (pooled, all timepoints)
0.9335
—
$R^2$ normalized (per-sample mean)
0.9219
[0.9214, 0.9224]
$R^2$ original scale (pooled)
0.9654
—
MAE normalized
0.0541
—
MSE normalized
0.00589
—
Per-curve $R^2$ (normalized): 0.936, 0.935, 0.933, 0.918, 0.931, 0.938, 0.936
Max-value prediction (curves 1–6)
Metric
Value
95% CI
Pooled $R^2$
0.9711
[0.9709, 0.9713]
Per-curve $R^2$ : 0.946, 0.946, 0.928, 0.907, 0.869, 0.783
Metric
Value
Active dims (var ≥ 0.01)
25 / 30
Collapsed dims
5 / 30
PCs for 90% / 95% / 99% var
20 / 22 / 23
Source
Mean
95% CI
Median
% > 0.9
% > 0.95
Real test data
0.9875
[0.9869, 0.9881]
0.9913
99.7%
97.8%
Generated (raw)
0.9588
[0.9560, 0.9615]
0.9868
85.7%
76.8%
Generated (θ=0.005)
0.9680
[0.9660, 0.9701]
0.9876
91.5%
80.5%
Novelty (nearest-neighbor distance, Euclidean on flattened curves)
Distance
Mean
Generated → nearest train
3.4295
Generated → nearest test
3.4402
Test → nearest train (baseline)
3.6235
Generated internal NN
3.8751
Memorization ratio (gen/train ÷ test/train)
0.946