Software Neurotechnology.
Twenty layers. Five categories. Zero weight changes.

Holds the live turn: prompt, tools, retrieved snippets, scratchpad.

Reaches back into prior turns, files, and sessions without retraining weights.

Anchors answers to source documents, citations, and evidence.

Compresses stale context so relevance beats recency, at bounded cost.

Breaks the ask into typed sub-goals the model can reason over.

Orders tool calls, retrievals, and writes into a runnable plan.

Grades the draft against the plan before returning it.

Rewrites the draft when critique flags gaps — bounded loop.

Keeps every output inside the organisation's AUP and DPDP posture.

Confidence tracks probability of being right — no hallucinated certainty.

Discloses when the model is out-of-scope or the frame doesn't fit.

Independent scoring against a public standard — the outermost regulator.

Tracks the person's role, goals, and preferences across the session.

Grounds what the model is looking at — objects, layout, spatial relations.

Live external state: web results, APIs, connectors, meeting audio.

Voice, room acoustics, camera feed — the multimodal edge.

Uses recent turns to sharpen the next answer — no weight update.

Learns from tool return codes, test runs, and connector responses.

Records what the user accepted, rejected, or edited — reshapes future runs.

Long-cycle improvement across a deployed instance and its logs.