Symbolic first
The adapter receives candidate parses from the OS. It does not invent forms that the OS rejects.
Model adapter
Neural models belong above the symbolic core.
The adapter layer resolves contextual ambiguity using a model, but the symbolic OS remains authoritative. This keeps the moat intact and prevents model shortcuts from leaking into the runtime.
Contract
L5 output plus context becomes a resolved parse.
{symbolic candidates + context}
->
{resolved parse + confidence + fallback}
Confidence-gated
If confidence is low, return multiple symbolic candidates rather than forcing a hallucinated resolution.
Model-neutral contract
Gemma may be the first target, but the adapter contract must work for future open or sovereign models.
Training data as product
The symbolic OS produces structured examples that can be licensed independently of hosted inference.