Intelligence should be organised around sparse events, local predictive computation, and multi-timescale memory — not around massive dense matrices trained once at enormous cost.
What we ruled out
Frontier 1 is the research engine. Its output is not a product — it is a set of validated architectural principles that Frontiers 2 and 3 apply in practice, using existing frontier LLMs as the reasoning layer.
The Aurora project is the vehicle. Aurora proposes a new class of AI architecture: one where the primary unit is a typed event (not a token), computation is distributed across sparse predictive microfields (not a global attention mechanism), memory is a four-tier living structure (not static weights), and energy is a first-class constraint (not an afterthought).
Six hypotheses (H1–H6) are preregistered and falsifiable. They span event-centric representation, sparse microfields, local multi-timescale learning, episodic-semantic memory, counterfactual settlement for planning, and federated schema exchange. The question is not whether Aurora will outperform frontier LLMs on general benchmarks. The question is whether the architecture can achieve comparable task performance on structured domain workflows at substantially lower energy, and whether it can learn continuously without catastrophic forgetting.
When a hypothesis passes in a controlled research environment, the relevant component is implemented in a Frontier 2 or Frontier 3 product — alongside an existing LLM, not replacing it. The product ships, generates revenue, and validates the hypothesis at scale with real users.
Aurora does not need to beat the frontier to be valuable. It needs to demonstrate that its components work. If they do, the cost and accessibility implications for AI deployment are substantial.
Open questions
Q01
H1 tests this directly: Aurora-M0 should achieve task success within 5 percentage points of a LoRA-tuned 8B baseline while using at least 40% less energy per successful task. The evaluation is preregistered. Target: Q3 2026.
Investigated by: aurora
Q02
H3 tests local neuromodulated learning with offline consolidation across 10 sequential task families. If retained accuracy stays above 85% with forgetting below 10 percentage points, the answer is the latter — and continuous local learning becomes a viable alternative to scheduled retraining.
Investigated by: aurora
Q03
H2 tests whether precision routing retains 95% of full-activation performance while activating 30% or fewer microfields. H4 tests whether episodic-semantic memory outperforms vector-RAG on stale-memory errors and exception recall. Together they test the core claim that structured memory beats longer context.
Investigated by: aurora
Currently open
Engineering spec and pre-model gym environments are in progress. H1 preregistered evaluation is the immediate priority. Preprint target: Q3 2026.
