Study 003 · Can conservative alignment improve calibration?

Daybreak Neural · 8 November 2026

Completed exploratory development study. At a budget of 24 original calibration trials, blended alignment plus updating reached 67.47% mean balanced accuracy, versus 67.47% for updating alone. The paired difference was +0.0067 percentage points, with a conditional 95% run-bootstrap interval of −0.67 to +0.61 points. This study did not demonstrate an added benefit from alignment on the primary comparison.

The conservative rule mitigated the previously observed eight-trial failure for participant 3, while worsening results for two other participants relative to full alignment. The hypothesis was motivated by Study 002 outcomes on these same people and sessions. Participants 5–9 remain reserved.

All 180 scores were independently recomputed from 40,185 prediction rows (893 distinct test trials), all 25 interval rows were reconciled, and all 31,255 predictions for the seven unchanged methods exactly reproduced Study 002.

Context and methods

Study 002 found a short-prefix failure of full covariance alignment. Here we test one fixed conservative rule: lambda = n / (n + 48), where n is the accepted calibration count, and R_blend = (1 - lambda) R_source + lambda R_prefix. Target covariance matrices are whitened using the symmetric inverse square root of this blend. The source tangent reference and scaler stay fixed.

The constant 48 is an engineering choice, not an optimized or calibrated prior. With 24 accepted trials the target estimate receives one-third of the covariance weight. No alternative strengths are tried in this experiment.

Primary question: does blended alignment plus the same supervised update outperform update-only at 24 original trials? Secondary comparisons include the full-alignment combination at eight and 24 trials. We retain every original Study 002 method unchanged and require exact replication of its saved predictions.

Data access: all accepted day-one task trials fit each person's models. The first day-two task run supplies the original chronological prefixes 0, 8, 16, 24 and 48. Artifact exclusions follow those boundaries. The same five later runs provide accepted test trials at every budget. No later recording enters a fit or shared preprocessing reference. However, analysts have already seen these sessions' results in prior studies: this is development evidence, not fresh validation.

Each prediction uses a four-second imagery epoch. Source artifact flags define an offline accepted-trial analysis. Neither a positive point estimate nor a conditional interval establishes live performance, calibration-time savings or scientific novelty.

Execution record. The 27 original code cells were executed top-to-bottom in a fresh Python namespace, with tables and three figures saved in this notebook. This environment blocks kernel sockets, so an in-process runner was used; the interactive Jupyter interface was not tested here. The final reference-check cell was then executed against the saved outputs. Independent scripts reconciled all scores, selected bootstrap intervals and unchanged Study 002 predictions. No model was refitted to write this interpretation.

Setup and portable rerun

Select Python 3 (ipykernel). The definitions are included below, so this notebook can run independently of the accompanying engine file. Install the bundle's pinned requirements.txt in a separate Python environment for the closest reproduction. Missing dependencies can also be installed with %pip install numpy scipy pandas scikit-learn mne pyriemann matplotlib threadpoolctl.

The first run may download eight public MATLAB files. Cached source files are checked against recorded SHA-256 hashes. Only participants 1–4 can be loaded. NEURAL_DATA_DIR optionally points to an existing cache; otherwise eeg_data is used. Results go to results beside the notebook's working directory. No account or API key is required.

Data: /workspace/scratch/b15a265229b0/tmp/eeg_data
Results: /workspace/scratch/b15a265229b0/conservative_alignment_study/results

Fixed protocol and implementation

The timestamp records when this design was fixed locally, before Study 003 outputs. It does not make the development study an external preregistration. Supervised updating uses the unchanged Study 002 weights: target fraction n/(n+48) and total fitting weight equal to the source sample count. This classifier weighting is separate from the covariance blend, even though both use the same numerical count rule.

Expand the definition cells to inspect the loader, fitting paths, blending and paired whole-run bootstrap. Source preprocessing and all seven reference methods remain unchanged.

Design fixed: 2026-11-08T20:12:59.558558+00:00
Primary: {'budget': 24, 'contrast': 'blended_combined minus update_only', 'endpoint': 'equal-participant mean balanced accuracy', 'unit': 'percentage points'}
Blend: n_accepted/(n_accepted+48)
Development participants: [1, 2, 3, 4]

Execute the fixed comparison

All methods use the same accepted later-run trials. Six adaptation variants have identical zero-budget predictions. Source hashes, covariance endpoints, finite features, access boundaries, supervised weights and convergence are checked during execution.

Participant 1/4: loading verified recordings and fitting fixed methods
Participant 2/4: loading verified recordings and fitting fixed methods
Participant 3/4: loading verified recordings and fitting fixed methods
Participant 4/4: loading verified recordings and fitting fixed methods
Study003 completed.

budget	0	8	16	24	48
subject
1	0	8	16	22	45
2	0	8	16	24	46
3	0	7	14	22	46
4	0	7	13	20	35

budget	0	8	16	24	48
subject
1	0.0	0.143	0.250	0.314	0.484
2	0.0	0.143	0.250	0.333	0.489
3	0.0	0.127	0.226	0.314	0.489
4	0.0	0.127	0.213	0.294	0.422

Accepted calibration counts by original budget:
Actual target covariance weight:

Results

Balanced accuracy is mean recall over four classes. Points at different budgets use the same later test trials. The curves are descriptive; they are not a search for the best budget. Participant-level plots show whether an average hides different responses.

Primary comparison at 24 original trials

The aggregate gives each participant equal weight. The unchanged context baselines remain useful references but do not define the primary contrast.

subject	1	2	3	4	Mean
CSP + LDA	74.86	55.43	74.55	60.99	66.46
Alignment only	78.09	52.50	73.75	58.98	65.83
Bandpower + LDA	59.76	35.17	60.87	51.95	51.94
Blended alignment only	80.52	52.09	79.56	61.56	68.43
Blended alignment + update	78.83	48.15	76.90	66.01	67.47
Full alignment + update	78.52	49.01	74.64	64.73	66.73
Frozen tangent decoder	77.39	50.47	80.43	55.74	66.00
Raw tangent + LR	78.27	50.47	79.98	55.74	66.11
Supervised update only	77.49	49.42	76.45	66.51	67.47

	subject	estimate_pp	ci_low_pp	ci_high_pp
5	1	1.345	0.467	2.202
9	2	-1.264	-2.938	0.000
13	3	0.446	0.000	1.415
17	4	-0.500	-2.240	1.064
21	mean	0.007	-0.674	0.607

Secondary comparisons and the previously observed short-prefix case

These comparisons were specified before Study 003 execution but were motivated by known Study 002 outcomes. They do not provide independent confirmation. Intervals resample whole test runs and are conditional on these people, fitted models, fixed prefixes and sessions; they do not cover future-day or population variability or correct for the development history.

	budget	contrast	estimate_pp	ci_low_pp	ci_high_pp
4	8	blended_combined_minus_combined	5.664	3.773	7.915
22	24	blended_combined_minus_combined	0.745	-0.694	2.305
23	24	blended_combined_minus_frozen	1.467	-1.754	4.204
24	24	blended_combined_minus_blended_alignment	-0.960	-3.459	1.253

	Participant 3, eight-trial budget (%)
Bandpower + LDA	60.87
CSP + LDA	74.55
Raw tangent + LR	79.98
Frozen tangent decoder	80.43
Alignment only	49.41
Supervised update only	79.53
Full alignment + update	52.00
Blended alignment only	79.13
Blended alignment + update	78.64

Takeaways

The primary comparison is near zero. The conservative alignment rule adds no demonstrated benefit over supervised updating alone at the fixed 24-trial budget. The interval does not prove equivalence or rule out effects on other people or days. Leaving out one development participant changes the mean difference from −0.44 to +0.43 points.

A known failure was mitigated, not universally prevented. At eight original trials, participant 3 improved from 52.00% with full alignment plus updating to 78.64% with blended alignment plus updating. Update-only reached 79.53%, and the frozen model reached 80.43%. Across the four people, blended versus full alignment plus updating improved by 5.66 points (conditional 95% interval 3.77 to 7.92), driven largely by participant 3. Participants 1 and 2 worsened by 2.11 and 5.49 points. Participant 2 still scored 43.50% with the blend, below its 50.47% frozen score. These results do not establish broad robustness.

Other budgets remain descriptive. At 48 original trials, update-only averaged 70.17%, versus 69.98% for the blend plus updating. Blended alignment without updating averaged 68.43% at budget 24; the prespecified comparison adding updating was −0.96 points, with an interval spanning zero. Selecting a new winner after inspecting these curves would be another development decision.

Research decision. Keep update-only as the simpler baseline and retain the frozen blend as a comparator. Before further tuning, use a locked protocol for a one-time evaluation on the reserved participants. Those participants can test transfer to new people within this dataset; a later independent dataset and prospective sessions would still be needed for broader claims. This study has not opened the reserved data or executed that next evaluation.

Business implication. We have an auditable evaluation pipeline and a concrete failure mode to investigate. We do not yet have evidence of a superior proprietary decoder or a reduction in real calibration time. The useful product question remains: can an adaptation procedure deliver reliable performance with fewer newly labeled trials?

Your rerun on another computer

1. Download and extract the Study 003 bundle into a new folder, separate from Study 002. Keep the original results; a rerun overwrites the extracted results folder.
2. Open Neural_Decoder_Experiment_003_Conservative_Alignment.ipynb in your working Jupyter installation and select Python 3 (ipykernel). Use the pinned requirements.txt with Python 3.12 for the closest match. If packages are missing, run %pip install -r requirements.txt in a temporary notebook cell, then restart the kernel.
3. Run all cells. The notebook downloads only participants 1–4 if its cache is missing; an initial download may take longer than the computation. Alternatively, set NEURAL_DATA_DIR to your existing eight-file cache before running.
4. Copy the final REFERENCE CHECK output, including any mismatches or warnings. Send that output back, together with any error traceback. If downloading fails, send the failing URL and error; do not substitute a different dataset.

This checks computational reproducibility in another environment. It is not validation on new people or recording days. Do not adjust parameters to improve agreement or open participants 5–9. Tiny numerical feature differences across platforms can occur; the check reports scores and exact predicted-label agreement separately.

Sources

• Dataset and trial timing
• Euclidean alignment: He & Wu
• MNE CSP documentation
• pyRiemann TangentSpace
• scikit-learn LogisticRegression

The complete bundle includes the protocol, source hashes, pinned packages, standalone code, predictions, access audit, matrices' diagnostic summaries, figures, reference results and independent validation.

REFERENCE CHECK
Score rows: 180 / 180 expected
All scores match within 1e-8: True
Maximum score/count difference: 0
Prediction rows: 40185 / 40185 expected
All predicted labels match exactly: True
PASS: computation reproduces the saved reference.