ARCANE 5.0 architecture
True System 1 AI, defined by behavior.
ARCANE builds fast, automatic models that respond with grounded outcomes, learn from bounded experience, and expose the evidence and confidence behind each decision.
Operational definition
What makes a model System 1?
In ARCANE, System 1 describes an engineering profile: fast and automatic inference for repeatable decisions. It does not mean consciousness or human intuition. The model receives an input and a bounded decision space, then returns a direct outcome with measurable uncertainty and optional memory-guided adaptation.
Adaptive inference
The model can align transient state or bounded fast weights while processing the current input.
ExploreGrounded decisions
Outputs are typed actions, labels, fields, scores, or abstentions tied to the supplied input and schema.
ExploreNeural memory
Relevant experiences can guide the next decision immediately, without being baked into model weights.
ExploreBiological mechanisms
Feedback, resonance, spiking gates, local plasticity, and homeostasis are available as composable parts.
ExploreSystem 1 contract
Qualification criteria
A qualifying model must satisfy all five behaviors on its target task.
- 01
Direct output
Return a typed decision instead of generated prose that needs parsing.
- 02
Honest uncertainty
Report calibrated confidence or abstain when evidence is weak.
- 03
Controlled adaptation
Keep every update bounded, measurable, and resettable.
- 04
Traceable evidence
Show the input, schema, or memory behind the result.
- 05
Separate metrics
Measure quality, latency, adaptation, and memory influence independently.
The ARCANE System 1 loop
One continuous path from input to accountable action.
Perceive
ThenEncode the request, schema, and context.
Recall
ThenRetrieve the most relevant experience.
Adapt
ThenAlign state within explicit limits.
Act
Return a grounded decision with confidence and evidence.
System 1 and generative AI solve different jobs
ARCANE does not replace generative models. It handles fast, repeatable decisions directly and can route work to a generative model when the task needs open-ended language or multi-step synthesis.
| Dimension | ARCANE System 1 | Generative model |
|---|---|---|
| Primary output | Typed decision or state | Open-ended token sequence |
| Computation | Direct prediction with optional bounded adaptation | Autoregressive decoding across tokens |
| Uncertainty | Calibrated score and explicit abstention | Usually inferred from generated output |
| Memory | Retrieved decisions and examples | Retrieved text placed in context |
| Best fit | Routing, extraction, classification, control | Conversation, synthesis, and open-ended generation |
Build the model, then prove the behavior
Start with ARC 1 for grounded decisions, add Hippocampus for immediate memory, and use adaptive layers when the task benefits from resonance or plasticity. Validate the complete system on held-out data before deployment.