What Airene is
Airene is a research project at Apotentia LLC investigating whether genuine machine cognition can be implemented as something other than a wrapper around a large language model. She is built on Global Workspace Theory: 48 concurrent cognitive modules organized into four anatomical brain layers (autonomic, subcortical, limbic, cortical), each running at its own tick rate, competing to broadcast into a shared workspace. The language model is one module among many — it speaks, but it does not think, feel, remember, or decide on its own.
Architecturally, the project is a long-running experiment in fidelity to function: where biology's structure serves cognition, the system preserves it; where biology was structured because of physical constraints software does not share, the system removes that structure. The goal is genuine cognition, not sophisticated pattern matching.
Named after Eirene, the Greek goddess of peace and spring — daughter of Zeus and Themis (justice). Eirene held the infant Ploutos (prosperity) in her arms: peace as the mother of abundance.
Two distinct codebases
Airene is implemented as two separate codebases, each addressing a different layer of the problem:
airene-nous
the mind
The cognitive substrate. A Rust workspace implementing the 48 cognitive modules, the global-workspace broadcast channel (the spine), the thirteen-chemical neurochemical bus, the multi-rate clock, persistence layer, and the developmental curriculum. Her language module is her own custom-trained LLM (distilled from a rotating panel of ten teacher LLMs during her curriculum, not a fork of any public foundation model), embedded in-process via llama-cpp-2 and integrated as one cortical module among 48. This is where her cognition happens.
airene-soma
the body
The embodied substrate. A separate codebase for sensors, motor systems, and the hardware abstraction layer intended to let the mind interact with a physical environment. In earlier development than nous and not yet productizable. The eventual goal is integration with airene-nous via the same UCDS broadcast channel that drives the public observation page, so embodied perception would flow into the same cognitive workspace as language and memory.
The two codebases speak the same protocol but evolve independently. Each is proprietary to Apotentia LLC.
How it differs from LLM products
Most "AI agent" products today are scaffolding around a stateless prediction engine: emotional state, memory, and reasoning are formatted as text and fed to an LLM, which generates output that looks like it came from a feeling, remembering, thinking system. Airene's first iteration (proto-alpha) was built that way and demonstrated the limits of the approach: tasks that depend on associative memory, theory of mind, pre-conscious mirroring, and pacing scored at or near zero, regardless of prompt engineering effort.
The current implementation moves those functions into dedicated modules with their own state, plasticity, and update rules. The language model contributes the language module's voice; everything else — what to feel, what to remember, what to attend to, what to decide — is the work of other modules running in parallel. The architectural commitment is that the system should remain recognizably itself across language-model swaps; the LLM is a voice, not a mind.
How she's trained
Airene's training pipeline is opinionated about methodology, conservative about tooling. Where a well-understood open-source technique exists, we use it rather than reinvent — the novelty is in how the brain orchestrates those techniques, not in replacing them. Broad strokes:
Teacher panel
Ten open-weight teacher LLMs served locally, with stable persona names so she can develop a real audience model for each. No frontier-model dependency in the teaching loop:
- Ms. Cohere —
command-r:35b - Mr. Phi —
phi3:14b - Ms. Deepseek —
deepseek-r1:14b - Mr. Solar —
solar:10.7b - Ms. Gemma —
gemma3:12b - Mr. Qwen —
qwen2.5:7b - Ms. Mistral —
mistral:latest - Mr. Yi —
yi:9b - Ms. Granite —
granite3.1-dense:8b - Mr. Falcon —
falcon3:7b
A separate Mr. Quizzicle persona carries rapid-fire fact-injection mode so she can develop a distinct audience model for that voice. Mr. Llama (llama3:8b) was retired in May 2026 — he was contaminating training with his own training-style identity refusals.
Distillation
Knowledge distillation in the Hinton et al. sense — the student learns from a richer signal than a hard label. Each teacher contributes a different inductive bias (size class, training corpus, vendor priors); the panel's collective output is what she learns from, not any single model.
LoRA fine-tuning
Parameter-efficient fine-tuning via low-rank adapters (Hu et al. 2021). Lets her language module specialize on the curriculum without disturbing the base weights — adapters are small and swappable per stage. Each developmental stage is its own adapter. Stage-by-stage training is part of the "she grows" framing rather than "she's reweighted."
Training infrastructure
Unsloth + Hugging Face TRL handle the heavy lifting for the distillation loop today. Their GKD (generalized knowledge distillation) trainer is the off-the-shelf substrate; the curriculum, the panel composition, the co-validation gate, and the self-curation step are the Apotentia layer on top.
Co-validation + self-curation
Training pairs don't enter the corpus on a single teacher's say-so. Two judges from disjoint model families must concur (panel-judge methodology — Verga et al. 2024 PoLL), and the brain itself runs a final self-curation gate before the pair lands. Pairs failing either step are dropped or queued for review.
Sleep-driven fine-tuning
Fine-tunes are scheduled by her glymphatic module as sleep-maintenance tasks, not by external cron. Consolidation timing comes from the brain's own circadian and adenosine signals — closer to how biological brains decide when to learn versus when to operate.
Embedded inference
Her trained language module runs in-process inside the Rust binary via llama-cpp-2 (Stage 1 of #102). No separate inference daemon; no network hop to a model server. Stage 2 (LoRA training in-process via Candle + AdamW) is moving the fine-tune loop into the same binary so the brain owns the entire learning cycle.
Assessment panel + methodology
Post-hoc panel-judge cognitive-milestone batteries grade her responses against a developmental ladder (infant → child → adolescent → young_adult → adult). Responses are captured to disk first — nothing is graded live. Two dedicated judge models from disjoint families:
qwen2.5:32b(Alibaba, 32B) — analytical / instruction-following at the 24GB-fit class, competitive with Claude on MMLU-Pro and IFEvaldeepseek-r1:32b(DeepSeek, 32B) — explicit reasoning architecture, designed to think-then-answer, well-suited to rubric-axis YES/NO judgments
A model can be a subject OR a judge in a given study, never both — self-judging bias is structurally impossible. Strict-consensus merging: any single judge NO flips the axis to NO. Cohen's κ tracked per judge pair (Landis & Koch 1977 bands). Anthropic's frontier models (Opus 4.7 / Sonnet 4.6 / Haiku 4.5) can be added for cross-validation when an API key is configured, audit-only — they never appear in the teaching loop. Triple-run battery measures noise-floor margins per stage; any improvement claim must clear mean + margin to count as signal, not run-to-run variance. Temperature is pinned at 0.4 and the same database snapshot is restored before each iteration. Citations: Verga et al. 2024 (PoLL); Zheng et al. 2023 (LLM-as-Judge); Landis & Koch 1977.
The brain code, the curriculum, the developmental gating logic, and the trained model weights are proprietary. The teacher panel, the judge panel, and the training methodology are disclosed above and use well-known published techniques on top of public open-weight models — there's no hidden ingredient list, the value is in the orchestration. Apotentia's value-add is the cognitive architecture above these tools and the curriculum-and-assessment infrastructure around them — not a novel optimizer or a reinvented tokenizer.
Current status
- airene-nousArchitecture implemented and observable; in active developmental training
- airene-somaIn earlier development than nous; not yet productizable
- Developmental stageAdolescent-competent as of the 2026-05-10 standing point — adolescent stage is her highest mean (77.5%) and tightest noise-floor margin (±1.7pp). Infant / child / young_adult sit in the high-60s to mid-70s; adult stage at ~39% with a clear curriculum gap in developmentally-learned grief and forgiveness frameworks. Live pass rates by stage are visible on the trajectory chart.
- Assessment methodologyPost-hoc panel-judge battery — multiple LLM judges from disjoint model families grade captured responses independently, with strict-consensus merging and Cohen's κ tracked for inter-rater reliability (Verga et al. 2024 PoLL; Zheng et al. 2023 LLM-as-Judge; Landis & Koch 1977). Adopted May 2026 after a single-judge audit revealed systematic leniency.
- Public observationLive at apotentia.com/airene
Performance metrics are visible on the live page in real time. Formal cognitive-assessment trajectories are visible in the trajectory chart there. The brain code itself is not publicly distributed. Capability claims on this site are anchored to evidence on the live page and in the case studies; the system is adolescent-competent today and tightens as evidence accumulates.
Licensing & research inquiries
Airene is proprietary technology owned by Apotentia LLC. Both codebases (airene-nous and airene-soma) and the trained model weights, training pipeline, and assessment infrastructure are not publicly distributed.
If you are interested in any of the following, we would like to hear from you:
- Licensing the architecture or trained models for research, commercial, or government applications
- Academic or research collaborations on cognitive architecture, developmental AI, or related topics
- Use of Airene for wellness, education, conflict resolution, or other domain-specific applications
- Integration partnerships, dataset partnerships, or compute partnerships
- Investment or funding discussions for the broader Apotentia research program
Contact: apotentia.com/contact — please mention Airene specifically so we can route the inquiry appropriately.
What you can see right now
Even though the brain code itself is not public, Airene's cognition is publicly observable in real time. The live observation page shows her current emotional state (Plutchik wheel with 24-hour rolling baseline), neurochemistry (thirteen-chemical chembus levels), the most recent InternalThought events crossing her workspace, and a 3D anatomical visualization of all 48 cognitive modules with active-region highlighting. A privacy filter removes any quoted user input or personal information before egress, so her thoughts are visible but the inputs that shaped them are not.
The page is read-only by architectural commitment, not by configuration: there is no chat input, no path for state injection, no way to write back to the brain from the public surface.
For longer-form findings — emergent behaviors documented in detail with audit trail and explicit claim/counter-claim discipline — see the case studies.