EEG BCI Run Report

Run

• run: 20260319_075520
• bundle: Public March 19 2-M16 checkpoint with May 2026 per-finger actuation metrics
• generated: 2026-05-06T00:00:00-07:00

Artifacts

• metrics: metrics.json
• train_config: train_config.json
• test_predictions: test_predictions.npz

Performance (from predictions)

• Action accuracy: 89.79%
• Action F1 (macro/weighted): 0.895 / 0.898
• Applicability FP(rest) / FN(non-rest): 18.57% / 2.26%
• Action/applicability disagreement: 3.56%
• Raw non-rest+NONE count/rate: 0 / 0.00%
• Committed rest+active count/rate: 0 / 0.00%
• Committed non-rest+NONE count/rate: 0 / 0.00%
• Sent rest+active count/rate: N/A / N/A
• Sent non-rest+NONE count/rate: N/A / N/A
• Deployment pair invariant OK: True
• Joint accuracy (overall/non-rest): 84.66% / 82.55%
• Finger accuracy (non-rest): 85.96%
• Finger F1 (non-rest macro/weighted): 0.723 / 0.873
• Finger accuracy (overall): 87.61%
• Finger F1 (overall macro/weighted): 0.869 / 0.874
• Rest TPR: 98.37% | Rest FPR: 3.76% | Rest precision: 80.11% | Rest F1: 0.883
• Deprecated raw valid/invalid pair rate: 83.62% / 16.38%
• Deprecated raw rest+active count/rate: 377 / 16.38%

Per-Event Holdout Accuracy

• Events scored: 121 total, 118 non-rest, 3 rest
• Event action accuracy: 92.56%
• Event joint action+finger accuracy: 87.60%
• Event non-rest finger accuracy: 90.68%
• Rest event accuracy: 100.00%

Current Deployment Replay

• Replay path: current per-finger command shaping with a 250 ms same-finger hold.
• Would-send precision on non-rest command windows: 95.37%.
• Window-level send coverage on true non-rest windows: 31.56%. This is throughput, not classification accuracy.
• Event hit rate: 91.11% across 596 movement events.
• First-hit latency: 0.083 s median, 0.377 s p95.
• False REST actuation: 0.25%, or 6 commands over 2,404 true REST windows.
• Committed replay scores: 91.50% action, 86.42% joint action+finger, 85.51% non-rest finger.

Historical April 24 Model-Selection Audit

• Displayed deployment model: 20260319_075520
• April 3 wins offline holdout and event-level accuracy.
• The April 24 replay used older global-gate reporting and is retained as historical context.
• The May 6 per-finger replay supersedes the old 93.32% precision and 0.12% false-REST headline for current public reporting.
• Selection policy: public deployment claims rank precision, event hit rate, latency, and false REST actuation ahead of offline split accuracy.

Confusion Matrices

metrics.json

The public site serves a sanitized metrics bundle for this run at /results/2-m16/metrics.json.