Conversation Brain

Conversation Brain is an Experimental native GPUI visualization of an interaction, not a claim that EXEPERT or a connected model has a biological brain. It gives Native Chat and Codex the same two-actor presentation:

• HUMAN · PREDICTED ANALOGY represents observable user interaction and an optional deterministic analogy inferred from visible English or Malay text.
• AI · INTERFACE ACTIVITY represents observable request, response, tool, approval, completion, and interruption lifecycle events. Only visible agent messages may receive optional text inference.

The two brains reuse EXEPERT's immutable low-LOD presentation geometry and have independent render caches. They do not mutate, subscribe to, or impersonate the existing Brain simulation worker. The compact card is rendered first in both Chat right panels; selecting its viewport opens the full BRAIN CONNECTION view. SIMULATION remains the original renderer experience.

Evidence model

Every readout is assigned one of four provenance classes:

Label	Meaning
OBSERVED	A submit, response, tool, approval, completion, cancellation, or failure occurred at the application interface.
INFERRED TEXT	A bounded local heuristic classified user-visible English or Malay text. It is not a hidden-state measurement.
HUMAN ANALOGY	Presentation mapping used to make the interaction legible. It is not measured neural activity.
ACTIVATION	Reserved for a future separately reviewed provider. The commercial core reports ACTIVATION · UNAVAILABLE.

The UI and manifest explicitly say that the display is not consciousness, hidden reasoning, literal thoughts, a measured human brain, clinical evidence, or discrete-emotion localization. Tone changes presentation color and activity intensity; no emotion is assigned to a single anatomical structure.

Data flow and exclusions

The runtime is process-local and retains at most 40 conversation states and 64 queued events. Native identities use the active local Chat ID. Codex identities use the server-qualified (server_id, thread_id) task key, preventing local and remote tasks with the same thread ID from colliding. Restart clears all derived activity. Only three per-device preferences are persisted: semantic inference, empathy mode, and collapsed-card state.

The following inputs are permitted:

• submitted Native or Codex user text;
• Native visible response deltas and final response text;
• authoritative Codex agentMessage deltas and completed items; and
• lifecycle metadata that an event occurred.

Raw reasoning, reasoning summaries, plans, command output, tool payload text, approval content, attachments, credentials, protocol payloads, and human-room messages never enter the classifier. A tool or approval can change the generic observed phase without exposing its content.

Semantic inference and empathy

Semantic inference defaults on. The bundled deterministic classifier bounds input to 4,000 characters, recognizes English and Malay markers, and projects six broad interaction channels: language, attention/executive, memory/retrieval, social-affective, visual, and auditory. Unsupported language or confidence below 0.45 produces no presentation guidance.

Empathy mode defaults off. When enabled, draft analysis runs locally after a 400 ms idle period; drafts are not sent before Send. The resulting tone and depth suggestion is visible and can be cycled by the user. Native Chat sends only validated enums and confidence. The Edge Function rejects unknown keys, free-form instructions, unsupported languages, and out-of-range confidence, then maps accepted enums to trusted fixed copy. The instruction is explicitly presentation-only and cannot weaken facts, uncertainty, safety, permissions, tool behavior, or the requested format.

For a new Codex turn, the same approved suggestion becomes a separate visible EXEPERT interaction cue text input. It is never added to steering, interrupts, approvals, or tools. A direct App Server client remains outside this EXEPERT presentation contract.

Geometry and research adapters

assets/brain/conversation-atlas-v1.json records the display geometry as reference-mapping-only. ICBM 152 and CerebrA are cited as coordinate-space and high-level anatomical references; EXEPERT does not bundle or claim to have registered their atlas data in this feature. The current geometry remains EXEPERT's original low-LOD presentation mesh.

BrainEvidenceProvider is the narrow future adapter seam for activation-based evidence. The shipped provider is deliberately unavailable. A future J-space, TRIBE, or other interpretability adapter requires its own licensing, model, layer, reproducibility, privacy, performance, and scientific-validity review before it can emit ACTIVATION evidence.

Verification boundary

Rust unit tests cover settings, bounded eviction, English/Malay determinism, guidance source-text exclusion, lifecycle direction, atlas disclosure, and the unavailable provider. GPUI source contracts lock stable IDs, both right-panel placements, dual canvases, provenance copy, Connection tabs, and Codex text exclusions. Edge validation has a standalone Deno test for enum acceptance and prompt-injection-shaped rejection. Live Windows rendering, keyboard traversal, reduced-motion behavior, and model response style remain manual checks.