DAYBREAK NEURAL · STUDY 007 · DEVELOPMENT PILOT

Does generated EEG help a small decoder?

Generated augmentation performed worse on average than real-only training at the fixed 24-label budget in this eight-person pilot.

Daybreak Neural · 9 November 2026 · 8 participants, 3 training seeds each

The result

At the fixed primary budget, the generated arm improved the seed-averaged score for 1 people, worsened it for 7, and tied for 0. All eight people and every seed are retained. These are OpenBMI participants 2–9, already examined in Studies 005 and 006. This pilot is exploratory, despite fixing its protocol before the new computation.

Balanced accuracy is the mean recall for right- and left-hand imagery. The 50% reference applies to uniform guessing in this two-class task. Each person has equal weight. The 52 real evaluation trials per person are reused across methods, budgets and seeds; 14,144 prediction rows represent 416 distinct test trials.

Every participant at 24 labels

What we actually tested

A compact EEGNet-style convolutional network learned from 100 real day-one trials. That fixed source checkpoint initialized each adaptation arm. The real day-two calibration budget was 8,16 or 24 chronological trials; tests always used day-two trials 49–100. The intervening trials were unused. No test labels or test waveforms fitted the generator, normalization, decoder or stopping rule.

• Real only: each adaptation batch contained 16 weighted real examples.
• Simple augmentation: the first 8 were identical real examples; the last 8 received mild amplitude scaling, Gaussian noise and occasional single-channel dropout.
• Generated augmentation: the first 8 were identical real examples; the last 8 were generated examples with the same class labels as the corresponding real draws.

The same batch size, number of optimization steps, source checkpoint and real-example sampling schedule were used in all arms. Fixed compute means that augmented examples replace some real presentations; the generator itself requires extra computation. This is not a claim of equal total computational cost.

The generator and its assumptions

A conditional variational autoencoder learned the distribution of source-standardized log covariance matrices from only the permitted source and calibration trials. It then sampled class-conditioned covariance matrices. Gaussian noise was spectrally shaped in 8–30 Hz, whitened, and colored to each sampled covariance to produce 62-channel EEG-like waveforms. The spectral envelope was estimated separately per class from permitted real training data.

This hybrid models spatial covariance and a restricted spectral envelope. It does not learn complete temporal dynamics, task timing or neural mechanisms. Calling a synthetic trial “left-hand imagery” supplies a conditioning label; it does not verify that the waveform contains valid left-hand neural evidence. There is no claim that this is a physiological simulator or that it tests all generative AI approaches.

The conditional VAE used a 64-unit hidden layer,12 latent dimensions, a 0.01 KL term,500 optimization steps and 128 generated trials per class. The EEGNet-style variant used 1,114 parameters,500 source steps and 100 adaptation steps. These CPU-bounded settings were fixed rather than optimized. Numerical quality checks passed for all 72 generator fits, but are insufficient to validate class semantics.

Preprocessing and classical comparison

All 62 EEG channels were processed at native 1 kHz with a causal fourth-order 8–30 Hz filter using a 2-second pre-cue context, then cropped to four post-cue seconds and demeaned. An explicit antialias resampling step produced 512 samples at 128 Hz. Per-channel RMS normalization used only the 100 source trials. No artifact-based exclusions were added. The source/target sampling fraction was 48/(n+48) versus n/(n+48).

A source-whitened OAS covariance, Riemannian tangent-space and logistic-regression pipeline was rebuilt on these same 128 Hz inputs. Its feature reference and scaler used source data only. It is a useful simpler comparator, but its scores are not the historical Study 005/006 scores: both the cohort and preprocessing differ. The EEGNet implementation is an independent compact variant, not an exact reproduction of the published default architecture.

Possible explanations, not established causes

The generated waveforms model covariance and a restricted spectral envelope, so they can omit task-relevant temporal structure. Their RMS amplitude was consistently below the overall permitted real-training RMS (ratios 0.81–0.92), which suggests a distribution difference but does not prove why performance fell. Half of each adaptation batch used synthetic replacements, reducing real-example presentations at the fixed step budget. Finally, the compact neural baseline itself trailed the classical reference. These are plausible contributors to test separately; this experiment does not identify a single cause.

Uncertainty and what this cannot establish

We averaged the three seeds within each participant before computing the primary mean paired difference and Student t interval with 7 degrees of freedom. The interval assumes independent, approximately normal participant differences. Eight consecutively selected, previously seen people from one dataset are a weak basis for population generalization. Seeds and synthetic examples do not increase the number of people. All additional contrasts and budget curves are descriptive and unadjusted for multiplicity.

Leaving out one person at a time changes the primary mean to a range of -2.78 to -1.88 pp. This sensitivity does not replace the full eight-person result. There is no prospective, clinical, real-time or new-person reliability claim. A null or unfavorable result would reject this configuration as a demonstrated improvement; it would not show that all generative augmentation fails. A positive result would still require independently collected or untouched confirmation.

Decision for the project

Generated augmentation performed worse on average than real-only training at the fixed 24-label budget in this eight-person pilot. At 24 labels, the classical reference reached 74.70% versus 63.51% for real-only EEGNet, a descriptive 11.20-point gap. Simple augmentation reached 63.41%. The immediate priority is a competitive real-data neural baseline before investing in a larger generator. Preserve this exact result and the simpler covariance baseline. Before scaling to a larger generative model, examine the seed variability, real-only decoder strength and class fidelity of generated signals on development data. Any revised settings must receive a new protocol and a separate evaluation; do not tune against this pilot’s test outcomes and relabel the same data as confirmation.

The business objective remains lower real calibration burden with dependable later-session performance. Generating examples is useful only if it moves that real-data trade-off. This study measures accuracy at fixed trial counts; it does not yet measure calibration time, integration cost, customer value or product readiness.

Verification and provenance

All 272 score rows were reconciled against 14,144 predictions. All 240 saved neural-network checkpoints reproduced their saved predicted labels on the real test trials. Maximum probability replay difference: 1.11 e-16. A same-environment retraining spot check repeated the first participant and first seed at the primary budget: the source network, generator and all three adapted networks matched every saved tensor exactly. This single chain is not a full-cohort retraining replication. The participant-level primary interval was independently recomputed. Source checksums, frozen implementation hashes, trial boundaries and source-only normalization were verified.

This is a computational audit and saved-model replay, not independent full-training replication by another team. The design was privately frozen at 2026-11-09T16:57:50.574338+00:00, before the new raw-file requests. It is not an external preregistration, and the participants had already been examined in earlier studies. Source hashes were measured by this project, not signed by the dataset publisher.

Protocol SHA 256: f9695c08f3bfe0a56f854118b25846eae42db60c8b48d07bb473e3e86fe2d76e

Reference environment: {"python": "3.12.14", "torch": "2.15.0+cpu", "numpy": "2.3.5", "scipy": "1.17.0", "pandas": "2.2.3", "scikit-learn": "1.8.0", "pyriemann": "0.12", "threadpoolctl": "3.6.0"}. The results notebook is executed in a fresh Python namespace using an in-process runner; the interactive Jupyter interface and Windows ARM 64 training are untested.

References and files

• Lawhern et al.: EEGNet architecture; authors’ implementation.
• Kingma and Welling: Auto-Encoding Variational Bayes.
• OpenBMI dataset publication.

The accompanying bundle contains the fixed protocol, implementation, full scores, per-trial predictions, training logs, fit audits, numerical generator diagnostics and results notebook. Separate epoch and model archives allow saved-model replay and fuller reproduction without repeating the raw downloads.