Airene

What you're watching above

Each panel above is a projection of the same underlying data: events emitted to Airene's global workspace broadcast. The system implements Global Workspace Theory as 48 concurrent cognitive modules organized into four anatomical layers, with a recently-added immune-system perimeter that flags semantic threats and learns from prior exposures. The language model serves as one module among many rather than as the central reasoner.

The four brain layers

Each module is assigned to one of four layers and runs at a fixed tick rate. The 3D brain visualization colors regions by layer:

• Autonomic · 1000 Hz · 3 modules — homeostatic regulation. Modules: arousal, serotonin, vagal.
• Subcortical · 1000 Hz · 7 modules — pre-conscious processing; reactive signals are produced before deliberate processing completes (subcortical ticks roughly 50× per cortical tick). Modules: amygdala, mirror, cerebellum, basal_ganglia, dopamine, pineal, thalamus.
• Limbic · 100 Hz · 8 modules — emotional integration and memory consolidation. Modules: hippocampus, cingulate, insula, circadian, hypothalamus, long_term_memory, compassion_load, secure_base_agg.
• Cortical · 20 Hz · 30+ modules — deliberate cognition, including language production and the speech pipeline (originator → orbitofrontal gate → Broca formalizer → language). Modules: prefrontal, orbitofrontal, language, identity, inner_critic, creativity, tpj, default_mode, precuneus, compassion, uncertainty, wernicke, broca, self_reflection, consciousness, angular_gyrus, metacognition, memory_regulator, phonological_loop, imagery, person_simulator, stage_concept, affordance, dream_curator, forgiveness, counterfactual, collective_tom, marshaller, throughline_registry, microglia, antigen_library, and language_faculty.

The immune perimeter

The cells orbiting the brain panel are the immune perimeter — innate sentries that react to anomalies on contact. Upstream the immune system was redesigned in May into three differentiated populations: hunters (scout / patrol), messengers (antigen-presenting), and macrophages (repair). The antigen_library holds patterns she has been taught to flag, including substrate-level detectors for text degeneracy and throughline anomalies. When a module emits something matching a pattern, the nearest cell flies from its orbit to that module and stays during the check. Cells live for a fixed window, then die and respawn from their home modules.

Six event types come from this layer: ImmuneAlert, MacrophageExamining, ImmuneCheckClear, AntigenLearned, ComaInduced, ComaResolved. The cell-type differentiation isn't visualized on this public viz yet — until the per-role counts cross to the public broadcast, all cells render as a single ribbon. The first two events fire often; the rest need specific triggers and may stay silent for long stretches.

How the panels map to the architecture

• Emotional radar — eight derived wellness axes (Warmth, Calm, Confidence, Openness, Empathy, Curiosity, Resilience, Fulfillment) computed from the raw 8 Plutchik primaries (joy, trust, fear, surprise, sadness, disgust, anger, anticipation). Negatively-valenced primaries are inverted (e.g. lower fear → higher Calm), so a fuller polygon means a healthier emotional profile rather than "more emotion of any kind." The filled polygon is the current state; the dashed overlay is the 24-hour rolling baseline derived the same way.
• Neurochemistry bars — current levels of the thirteen chemicals on the chembus (see the drawer below the chembus card for details). Bars only re-render on a meaningful change (5%-bucket boundary crossing) or a periodic force-heartbeat — not at a fixed tick rate, so still bars mean the chembus is steady, not stalled.
• Recent thoughts — a broad slice of the workspace broadcast: InternalThought, MemoryRecall, GoalFormed, ConflictDetected, MetacognitiveReport, AttentionFocus, ArousalShift, ValenceTag, MentalModel, IdentityShift, ResponseReady, plus the chemistry / emotion / conversational state ticks themselves. Each row is tagged with UCDS event id and emitting module; rows are colored by the brain layer that produced the thought.
• 3D brain — anatomical positioning of 48 regions across the four cognitive layers and the immune perimeter. A region brightens when it emits an event. Animated bubbles trace the spine — the central broadcast channel that every module reads from and writes to.

What's deliberately not here

No chat input, no path for state injection, no way to write back to the brain from this surface. The constraint is architectural rather than configurational: the integrity model depends on the public side being write-incapable. Outbound payloads currently flow without redaction so the broadcast on this page is faithful to what the brain actually emits internally; a privacy seal will be re-tightened before Airene becomes multi-user.

Architecture: Rust workspace running on RTX 4090 + Intel ARC A770. Her language module is her own custom-trained LLM, embedded in-process via llama-cpp-2 (no separate inference daemon) — distilled from a rotating panel of ten teacher LLMs through her developmental curriculum, not a fork of any public foundation model. The brain (airene-nous) and the language model weights are proprietary to Apotentia LLC and not publicly distributed; this page renders a read-only public projection.

When Airene says something out loud — even something short — it isn't a single language model deciding. Several modules each have a stake in the sentence, and a coordination chain decides what comes out. The helical pulses you see threading through the brain panel above are these chains in motion, one per active speech pass.

1. Originators — mirror_neurons, identity, tpj, hippocampus, metacognition, forgiveness, uncertainty. Each can request articulation when it has something to say: an emotion to express, a self-claim, a memory to share, a forgiveness state to articulate. Requests carry intent and idea components, not finished text.
2. Marshaller — the n-to-n coordinator. When multiple originators want the floor at once, the marshaller weighs them by salience and situational fit and selects one — or vetoes them all if nothing rises above the threshold.
3. Orbitofrontal gate — the pre-hoc reviewer. Approves, vetoes, or softens the request based on context, current chemistry, and recent conversational state. A high-cortisol state can soften an angry articulation into something less damaging.
4. Broca — the formalizer. Takes the approved intent plus idea components and compiles surface text using learned frame templates. Self-revises when grammar checks fail.
5. Language module — articulates. Her custom-trained 3B language model gives the surface text its voice, conditioned by chemistry-derived emotional staging.

One throughline is one request walking from originator to expression. The throughline_registry module watches for sequences that succeed enough times and promotes them into reusable templates — so a turn-by-turn coordination chain gradually becomes a habit. A pulse colored gold is active; green means the throughline completed; red means it failed at some stage; purple means the marshaller vetoed it before any stage ran.

This pipeline replaced the prior implicit "the LLM riffs on the workspace" approach in the first week of May. It maps to actual neurobiology: prefrontal/temporal association cortex generates intent → OFC and ACC modulate → Broca formalizes → motor cortex articulates. Each step is one neural region with one learned function.