Daybreak Neural • Research record • Study 005

Does a small calibration set help on a new recording day?

Supervised calibration improved mean decoding accuracy on this independent dataset.

Completed fixed-cohort OpenBMI evaluation · 53 reserved participants · two-class offline motor imagery · 8 November 2026

Primary mean change+6.38 pp24 labeled calibration trials

95% interval+3.76 to +8.99Participant t interval, df = 52

People who improved38 / 5314 worsened; 1 tied

What we learned. At 24 calibration trials, frozen balanced accuracy averaged 65.50%, supervised updating 71.87%, and updating plus blending 72.03%. The primary paired gain was +6.38 percentage points (95% participant t interval +3.76 to +8.99); 38 participants improved, 14 worsened and 1 tied. The prespecified interval excludes zero in the positive direction. This supports an average benefit under this protocol; it does not guarantee a benefit for each person.

What was tested

Each person had a decoder trained using their own 100 day-one offline trials. On day two, the supervised update received the labels for the first 24 trials. Both the updated and frozen models then predicted the same later 52 trials (49–100). Trials 25–48 were unused at the primary budget. A prediction distinguished imagined right-hand from left-hand movement.

This tests a repeat-user calibration strategy on an independent dataset. It does not test a universal pretrained model transferring without user-specific training. Participant 1 was used only for structural and implementation checks; participants 2–54 were fixed before their recordings were requested.

Per-participant gains and losses from supervised updating, with the mean and 95 percent interval — Balanced accuracy averages recall for the two classes, so unequal class counts in the test block do not dominate the score. Each participant receives equal weight in the primary result.

Calibration budgets and the blended method

	Frozen day-one decoder	Supervised update	Update + covariance blend
0	65.5	65.50	65.50
8	65.5	68.68	69.48
16	65.5	70.27	70.41
24	65.5	71.87	72.03
48	65.5	71.42	72.11

Table entries are mean participant balanced accuracy (%). The primary budget is 24; all other budgets are descriptive. The frozen decoder uses no day-two calibration at any listed budget.

Mean balanced accuracy by calibration budget for all three methods

At 24 trials, adding the covariance blend changed mean performance by +0.15 percentage points relative to supervised updating alone. This is a secondary descriptive comparison, with no confirmatory superiority claim. We did not select a best budget or tune the blend after reviewing the reserved data.

How the decoder works

For each trial, 62 EEG channels at 1,000 Hz enter a causal 8–30 Hz filter with two seconds of preceding context. The next four seconds are centered and summarized with an OAS covariance matrix. A source-fitted whitening transform, Riemannian tangent map, scaling and logistic regression convert that matrix into a class prediction.

The update refits logistic regression on source trials plus the allowed labeled prefix. At budget n, target trials have combined relative weight n/(n+48), with total model-fit weight held at 100. The blended arm also mixes source and calibration covariance references with the same fixed fraction. All fitted references and scales exclude the test block.

Evidence and verification

All 53 participants completed; no participant was dropped.
795 score rows and 41,340 prediction rows cover 2,756 distinct test trials.
Every accuracy and balanced-accuracy score was independently recomputed from saved labels and predictions.
Fit-access records contain only source trials and the allowed chronological calibration prefix. The preflight label-permutation check left all predictions unchanged.
At zero calibration, all methods match exactly. Frozen predictions stay identical at every budget.
The protocol and implementation hashes match the saved pre-data freeze. All 106 first recording requests occurred after that freeze.
63.63 GB of source files were processed. Their measured hashes and URLs are recorded; the raw EEG files are excluded from this report bundle.

What this means for the business

The product hypothesis is that BCI teams can preserve decoding performance across recording days while collecting a limited amount of labeled calibration data. This study measures one part of that hypothesis: held-out decoding performance under a fixed offline protocol. A benefit over a frozen decoder would support the value of calibration; it would not establish a proprietary advantage over other calibration methods or products.

The positive mean result supports continuing toward a partner-data pilot with the simple supervised update as a baseline. A useful next test would measure repeated-day performance and actual calibration time on a partner’s hardware, while monitoring users whose accuracy worsens. A safe update-selection or rollback rule would require its own development data and a new untouched evaluation set.

Limits on the conclusion

Public data from healthy participants, one fixed day pair, offline two-class motor imagery. Clinical users, other tasks and online feedback may behave differently.
Only 52 later trials per person; individual differences are noisy. The interval estimates a mean under an independent-participant sampling model and is not a prediction interval for a new user.
No physiological artifact rejection. Structural and numerical checks do not establish that all EEG is artifact-free.
Trial budgets are counts, not guaranteed minutes saved; pauses and inter-trial timing vary.
All reserved OpenBMI participants are now evaluated. Any tuning prompted by these results needs a new evaluation cohort.
This was a private prospective freeze, not externally registered preregistration or peer review. Prior BNCI results and OpenBMI participant 1 informed the design.

Participant-level primary results

All values are percentages except the final paired difference in percentage points.

	Frozen day-one decoder	Supervised update	Update + covariance blend	Update gain (pp)
subject
2	61.54	90.38	90.38	28.85
3	89.39	88.33	90.00	-1.06
4	60.45	72.27	67.73	11.82
5	84.70	83.03	82.42	-1.67
6	68.18	70.91	69.24	2.73
7	63.79	70.76	70.76	6.97
8	49.55	56.97	59.70	7.42
9	56.06	58.64	58.64	2.58
10	62.73	60.61	67.88	-2.12
11	51.67	75.91	70.76	24.24
12	53.33	67.58	72.58	14.24
13	52.27	74.70	76.36	22.42
14	50.00	53.18	53.18	3.18
15	43.79	71.52	65.45	27.73
16	47.27	47.58	44.24	0.30
17	59.62	57.69	59.62	-1.92
18	78.18	84.85	84.85	6.67
19	64.24	68.94	66.67	4.70
20	70.30	65.61	62.27	-4.70
21	97.73	96.67	98.33	-1.06
22	55.76	64.09	67.42	8.33
23	75.45	75.91	75.91	0.45
24	53.48	55.26	56.96	1.78
25	52.73	82.12	81.52	29.39
26	48.79	65.45	64.85	16.67
27	41.36	50.45	50.45	9.09
28	81.82	93.18	93.18	11.36
29	95.00	98.33	98.33	3.33
30	74.24	70.00	66.67	-4.24
31	72.12	77.12	80.45	5.00
32	70.45	91.06	87.73	20.61
33	98.33	96.06	97.73	-2.27
34	49.09	52.73	56.06	3.64
35	78.18	80.00	78.33	1.82
36	100.00	100.00	100.00	0.00
37	93.33	96.67	96.67	3.33
38	60.29	62.37	59.22	2.07
39	81.06	76.36	76.36	-4.70
40	53.33	57.12	57.12	3.79
41	60.00	53.64	53.64	-6.36
42	54.09	62.27	64.55	8.18
43	66.06	62.12	63.79	-3.94
44	75.00	93.18	93.18	18.18
45	90.38	94.23	92.31	3.85
46	63.94	74.85	72.58	10.91
47	44.55	70.00	75.61	25.45
48	63.18	65.76	68.03	2.58
49	58.79	74.39	76.06	15.61
50	50.00	58.03	54.09	8.03
51	60.91	63.33	61.67	2.42
52	64.39	61.97	65.30	-2.42
53	62.88	61.97	64.70	-0.91
54	57.58	53.18	55.91	-4.39

Sources and reproducibility

OpenBMI data deposit · Original dataset paper. The companion bundle includes the exact protocol, model code, acquisition record, per-trial predictions, validation script and executed results notebook. Rebuilding this report does not refit the EEG models.

Protocol frozen: 2026-11-08T22:09:08.192068+00:00
Protocol SHA256: fc55100319d2259a7ef7469c1eb812fa76065d3b9beb03e5ba59e3c0c0cf3e4b