01
Event-first substrate
Sensory input becomes sparse typed events, not flattened tokens. The system reasons over what changed, not over a window of text.
← All projects
Research projectModel Architecture
Aurora
A new class of AI architecture organised around sparse events, local predictive microfields, and multi-timescale memory — energy-bounded by design. Not a transformer. Not a fine-tune. Not a wrapper.
Relationship
Aurora is a standalone architecture research project, independent of any LLM vendor. When Aurora's hypotheses pass in controlled environments, the relevant components are implemented in Frontier 2 and 3 products alongside existing frontier LLMs — not replacing them.
Features
What it does.
02
Predictive microfields
Many small recurrent circuits, each predicting one local domain — objects, causality, language, time, risk. Errors are local. There is no global backpropagation.
03
Thalamic precision routing
A precision router gates which microfields fire based on uncertainty, task relevance, cost, and urgency. Computation goes where it is justified, nowhere else.
04
Four-tier memory
Working memory, episodic index, semantic schema lattice, and procedural memory — four distinct systems that replace the role of static model weights as the home for knowledge.
05
Offline consolidation
Sleep-like cycles replay episodes, test counterfactuals, integrate new knowledge, and prune low-utility memories — without user interaction.
06
Energy-aware objective
Reasoning depth, memory retention, and computation are explicit costs in the objective function. The system asks "is it worth reasoning more?" before it asks "what is the answer?".
07
Falsifiable hypotheses
Six preregistered hypotheses (H1–H6) span event-centric representation, sparse microfields, local learning, episodic-semantic memory, counterfactual planning, and federated schema exchange.
08
Federated schema exchange
Communities can share compressed schemas — causal routines, domain rules, affordances — without sharing raw personal data. Local intelligence with collective benefit.
Get started
Get started.
Aurora is research-grade. Read the concept paper and the preregistered methodology before engaging with the engineering spec.
Aurora is Anthril’s model architecture research project. It proposes a fundamentally different class of AI — one that treats the world as a stream of typed events rather than a token sequence, distributes computation across sparse predictive microfields rather than a global attention mechanism, and treats energy as a first-class constraint rather than an afterthought.
The name stands for Adaptive Unified Resonance Organism for Relational Autonomy. The architecture is inspired by neuroscience — not as a superficial metaphor, but as a source of engineering principles: sparse coding, predictive processing, multi-timescale memory consolidation, local Hebbian learning, and energy-regulated action selection.
What it is not
- Not a chatbot
- Not a transformer variant or token predictor
- Not a fine-tuned foundation model
- Not a RAG layer over a vector store
- Not a wrapper around an existing LLM
Six research hypotheses
Aurora’s development is organised around six preregistered, falsifiable hypotheses:
- H1 — Event-first representation achieves comparable task success to LoRA-tuned 8B baselines while using ≥40% less energy per successful task.
- H2 — Thalamic precision routing activates ≤30% of microfields per decision while retaining ≥95% of full-activation task success.
- H3 — Local neuromodulated learning with offline consolidation retains ≥85% accuracy across 10 sequential task families with ≤10pp forgetting.
- H4 — Episodic-semantic memory produces ≥20% fewer stale-memory errors and ≥15% higher exception recall than a vector-RAG baseline.
- H5 — Counterfactual settlement achieves ≥15% higher success per 1,000 environment interactions than model-free RL.
- H6 — Federated schema exchange improves cold-start task success by ≥20% relative to isolated local learning.
Aurora does not need to outperform frontier LLMs on general benchmarks to validate the thesis. It needs to demonstrate that its components achieve comparable task performance on structured domain workflows at substantially lower energy — and that they can learn continuously without catastrophic forgetting.
Status & roadmap
- 2026-Q1 · Neuroscience foundations and mathematical principles documented
- 2026-Q2 · Engineering spec, pre-model gym environments, and ADRs in progress
- 2026-Q3 · H1 preregistered evaluation (target)
- 2026-Q4 · H2 and H3 evaluations (target)
- 2027-Q1 · Public preprint (target)