Featured run
Current source: combined_20260319_081200_pruned_rest_events_0_1_2 · 20260319_075520
Action accuracy: 89.79%
Finger accuracy on non-REST windows: 87.01%
Current featured public bundle with March 19 metrics, figures, and replay summaries.
These sections are separated so saved split metrics, holdout diagnostics, and replay summaries are not presented as the same thing.
Saved test split
89.79% / 87.01%
Action and finger accuracy
Action accuracy is reported on all test windows; finger accuracy is reported on non-REST windows.
2,301 test windows; 1,994 non-REST
Primary holdout
84.66%
Joint accuracy with REST and applicability diagnostics
Adds REST and applicability diagnostics beyond the saved test split.
Deployment pair invariant passed; 2.26% applicability FN on true non-REST
Quiet-rest replay
97.26%
Auxiliary quiet-REST check
Separate replay used to characterize REST behavior.
4.53% applicability FP on true REST
Chronological replay
84.30%
Core full-session replay
Replay across the two core movement sessions on a longer chronological trace.
95.99% REST TPR; 18.07% applicability FP on true REST
Pseudo-live replay
86.64%
Saved Step 7 decision path
Replay through the saved decision path, including would-send precision and false REST actuation.
0.12% false REST actuation; 93.32% would-send precision
Harder replay session
71.96%
March 17 realism session
A harder pseudo-live check that exposes where applicability recall is still weak.
52.98% applicability FN on true non-REST
Tip of the iceberg
The featured 2-M16 run was not chosen on a single accuracy number. It emerged from repeated retraining, postprocess ablations, holdout audits, chronological replay, and pseudo-live replay until the deployment pair invariants stayed clean while the broader replay ladder remained competitive.
2,595 configs
Postprocess ablation
The March 16, 2026 website update documents a 2,595-config ablation over thresholds, smoothing, hysteresis, adjacency, and finger-mode settings.
96 retained sweep runs
Archived Step 2 + Step 3 cycle
The preserved `logs/sweep/` CSVs retain 96 completed training-plus-evaluation runs from the broader 2-M16 tuning cycle.
100+ model variants
Documented in the older Feb 26 update
The February 26, 2026 2-M16 tuning update states that 100+ model variants were trained across full-dataset, non-REST event-gated, and REST event-gated regimes.
30+ hours
Continuous sweep time
That same February 26, 2026 update describes a 30+ hour sweep and highlights a 90-run block that spanned about 33.3 hours from February 25, 2026 07:49 to February 26, 2026 17:07.
How the featured run was chosen
Training snapshot
Architecture
CNNLSTMFingerActionNet
The March 19 winning run combines an action head, an active-finger head, and a dedicated finger-applicability head.
Optimization
60 epochs · batch 64 · lr 0.001 · seed 43
These values come from the winning run's training config and match the published March 19 metrics bundle.
Split policy
group_trial · test_size 0.2 · calibration_size 0.1
The holdout bundle stays tied to a fixed split while calibration is separated from the main train/test partition.
Input + preprocessing
64 x 4 windows · center_detrend
Per-window centering and detrending are frozen into the winning run's preprocessing and normalizer config.
Interactive latent space
Each point is one EEG window projected from the learned latent representation into a three-component PCA view. PCA is shown here because it preserves a linear comparison across the full dataset, the training split, and the held-out test split.
The train/test pair stays colored by true finger so the geometry is directly comparable across splits. Correctness and deployment-gating views are available on the visualization page.
PCA · Full dataset
Each point is one EEG window, projected into three principal components and colored by the labeled finger.
Separated regions suggest structured learned organization, while overlap marks similar or harder windows.
Train/test comparison at a glance
Full dataset: 12,447 windows. Train split: 10,146 windows. Held-out test split: 2,301 windows. Scan the pair below for whether the held-out geometry still resembles the training structure under the same color coding.
Same projection family, same label colors, different split membership.
PCA · Train split
This restricts the PCA view to the 10,146 training windows while keeping the same true-finger coloring used in the full-dataset view.
Keeping train and test on the same coloring makes split-to-split geometry easier to compare directly.
PCA · Test split
This shows the 2,301 held-out test windows only, using the same true-finger coloring so the class structure can be compared against the training split.
If the held-out view preserves similar neighborhoods rather than collapsing, the representation is carrying structure beyond the fitting set.
These figures show error structure and confidence behavior for the featured run.
Action confusion matrix
REST, OPEN, and CLOSE confusion for the featured bundle.
Finger confusion matrix (non-REST)
Finger-level confusion after REST windows are removed from the task.
Calibration
Confidence versus observed accuracy for the featured bundle.
Confidence and uncertainty scatter
Confidence spread for the featured bundle.
These figures are signal-evidence context. They are not substitutes for the holdout and replay metrics above.
Action alpha rest-delta topomap
Rest-relative alpha maps for REST, OPEN, and CLOSE in the March 19 winning session. OPEN and CLOSE both show the dominant TP10 decrease and smaller TP9 increase that characterize the current 2-M16 action story.
Finger alpha rest-delta topomap with NONE reference
Finger-level rest-delta maps, including NONE as the explicit REST reference. The strongest variation remains concentrated on TP10 and TP9, which helps explain why lateral Muse 2 channels carry most of the finger-separation load.
The featured bundle is shown in context. The 1-M16 bundle is an earlier baseline and should not be treated as a direct comparison to the current 2-M16 run.
| Run | Role | Action accuracy | Finger accuracy on non-REST | Test windows | Links |
|---|---|---|---|---|---|
2-m16 March 19, 2026 | Featured run | 89.79% | 87.01% | 2,301 | |
1-m16-500 March 5, 2026 | Historical baseline | 83.94% | 80.61% | 2,652 |
Deep-dive page
The guided synthesis page for tuning scale, model choice, and recommended reading.
Featured run detail
The full training, tuning, replay, and selection story behind the current bundle.
Featured HTML report
The reviewer-facing report for the currently featured run.
Sanitized metrics JSON
Machine-readable metrics, figures, and benchmark fields for the current bundle.
March 19 winning-model update
The update that promoted the current deployment candidate onto the site.
March 16 split-fix update
The update that documents the 2,595-config postprocess ablation.
Older tuning update
The earlier update that documented 100+ trained variants and a 30+ hour sweep.
All dated updates
The chronological project trail behind the current metrics.
Paper status
Manuscript status page and public placeholder PDF while the pre-print remains private.