Research Note 000 · 2026-03-11
Information retrieval with neural signals.
morphism turns EEG into semantic embeddings and retrieves documents the wearer’s cognitive state points to. No language is used — the query is neural. The representations are coarse and approximate; what they track is the wearer, not the conversation around them.
The result
| Evaluation | n | Z-score |
|---|---|---|
| Signal chats — the wearer’s own messages | 442 | 5.34σ |
| Signal chats — the other person’s messages | 525 | 0.21σ |
| Claude chats | 246 | 10.78σ |
The dissociation is the finding. Same conversations, same recording — only the wearer’s messages decode. The model is reading the wearer’s cognition, not conversational structure.
How it works
Two stages: an autoencoder puts the EEG into a 64-dimensional latent space, then a semantic mapper guesses a semantic vector from the neural one. The predicted embeddings carry 62 dimensions of meaningful variance (at 95%) of the 1024-dimensional target — nearly everything the bottleneck preserves.
Get it
Weights and code are public under MIT at huggingface.co/evelabs/morphism. A 16-channel OpenBCI Cyton is required. Research preview: the output is associative, not transcriptive.