Research — Telepathy Codec

The research track explores decoding linguistic intent from neural speech pathways. We keep the downstream SPG + Fabric stable while swapping upstream inputs as research matures.

Conceptual Layers

1. N0 Signal IO — EEG/ECoG/MEG (research contexts); simulate with voice/gesture now
2. N1 Feature Bank — filterbanks, time-frequency patches, envelopes
3. N2 Phoneme Head — neural → phoneme posteriors over time
4. N3 Language Head — phoneme → subword tokens with LM prior
5. N4 Semantic Head — tokens → intents/constraints for SPG

flowchart LR
  Sensor[Signal Source (voice/BMI)] --> Feat[N1 Feature Bank]
  Feat --> Phon[N2 Phoneme Head]
  Phon --> Lang[N3 Language Head]
  Lang --> Sem[N4 Semantic Head]
  Sem --> SPG[Short Prompt Generator] --> Fabric[AGI Fabric]

Why staged?

• Ship Thought→Code now via typed/voice inputs.
• Mature the decoder separately; downstream APIs remain stable.
• Safety and privacy are easier to reason about with SPG contracts.