Native_Language_Model

Native Language Model

Orrin has two complementary language layers that are deliberately not the external LLM tool: a symbolic phrasing tier and a from-scratch neural language organ. Both learn only from what the runtime itself has read and said.

The two tiers

• Symbolic phrasing tier (brain/cognition/language_acquisition.py) — mines reusable discourse-marker openers from what Orrin reads and lends them to the template speech builder. It broadens how the template pipeline phrases things today, with no neural network involved.
• Neural language organ (brain/cognition/language/) — a from-scratch transformer that learns to produce language:
  • tokenizer.py — the tokenizer, built from Orrin's own corpus.
  • native_lm.py — the model itself (checkpointed as native_lm.pt).
  • acquisition.py (+ acquisition_noise.py) — the training loop over what the runtime reads; corpus hygiene matters here (a contamination fix keeps non-language artifacts out of training).
  • voice.py — the speech handoff: the organ is meant to gradually take over speech from the symbolic templates, gated so it only speaks where its output quality clears the bar.
  • conditional_render.py, library.py — rendering support and the phrase library.

Design intent

The point is developmental, and it's the same point that keeps the external LLM off the control loop: a pretrained model is frozen, and Orrin wants a language faculty that grows. An external LLM's weights never change with what this runtime lives through; the native organ's do. So rather than borrow a fixed model, Orrin grows his own — a transformer that trains only on what he has read and said, and gets better as that corpus of lived experience accumulates. The external LLM (see LLM Integration) never writes Orrin's inner speech; the native organ and the symbolic templates do, and the UI's Language panel plus the bilingual thought line (frontend/src/lib/thoughts.ts / lexicon.ts) let you watch the vocabulary develop. This is why the LLM is contained to a gated tool rather than made the controller — see Symbolic-First Design for the full argument.

Relationship to memory

memory/lexicon/ tracks learned word associations on the memory side; the language organ trains on the reading/speech corpus. The two develop together but are separate stores.

Practical notes

• Everything here is symbolic-only-safe: no provider key is ever required.
• The native_lm.pt checkpoint is runtime state; tests isolate it so a test run can't clobber a developing model.

Code pointers

• brain/cognition/language/ — the neural organ
• brain/cognition/language_acquisition.py — the symbolic phrasing tier
• frontend/src/components/brain/LanguagePanel.tsx — the UI surface