On this page
The thesis
Useful intelligence should be something we can keep building on.
ERAIS investigates how a system can accumulate knowledge, senses, tools and memory while selecting the work each task needs. Fracture brings supported trained components into that architecture; native training develops capabilities within it.
The practical question is whether this gives researchers and model builders a useful way to extend their systems within the compute they can afford. We measure retained behaviour, new capability, execution cost and the work required to integrate an upgrade.
Have a model, task or question you want to test? Tell us what you’re trying to find out. Bring the comparison that matters to you and the resources available for testing it. See how a research partnership works.
Current engineering
The 20 September engineering review checks running male and female CNS source models, complete native-network execution in Doom, and exact-state execution improvements. Its downloadable record names the source revision and measurement receipts.
The articulated-body comparison recorded 30.3% lower body CPU time on the measured laptop. Two retained-image neural-kernel comparisons recorded 6.2–7.1% lower CPU time, with complete mutable state and weights matching byte-for-byte. These are component measurements; the complete articulated simulation remains below real time on that machine.
Execution maps and source-bound sensory views let us inspect which components ran and which images a neural interval consumed. Calibrated body mappings, visual coverage and reliable long-running behaviour remain active work.
Earlier research results
The research snapshot published on 4 September 2026 records three completed paths:
- Native ERAIS: an 87.3M raw-byte model trained for 2,048 updates with adaptive patching and K1–2 sparse routing active.
- Fracture: learned speech and language-model components have been converted into real expert structures. In the public speech slice, the full eight-expert bank reproduced the donor waveform exactly; K4 and K2 retained zero proxy word edits on the 12-word evaluation phrase.
- Multimodal cognition: a real 1920×1080 source traversed the vision-to-world path, with 576 of 129,600 visual units entering expensive cognition.
These tests established concrete starting points. Their numbers describe those configurations; each later training run and deployment has its own receipt.
The implemented architecture extends beyond that snapshot. Fracture covers learned components and complete supported game-control policies. The public runtime update recorded on 11 September connected native text attention and KV execution with full-K converted language MLPs, while retaining donor tokenizer, embeddings and multimodal wrapping. Unified World, the Fracture API and Discord use the shared resident language service; Minecraft uses its own perception-and-action runtime.
Each checkpoint needs its own evaluation. The programme tests complete model conversion, sparse quality recovery, temporal reuse, capability-preserving growth and continuous multimodal operation on named profiles. The records distinguish implemented execution from trained and evaluated behaviour.
The questions that drive the programme
Does the learning hold beyond the training view?
Test it on unfamiliar examples and changed views. The useful result is behaviour that still holds when the task moves beyond the material used to train it.
Can routine detail be compressed without losing meaning?
Compare the compressed view with the complete input. Check that the order still makes sense and that an unexpected detail survives when it matters to the task.
Can stored capability grow without making each task more expensive?
Track quality and work per task as the stored capabilities grow. A larger system needs to earn its place through useful gains within the available compute budget.
Does any advantage survive a fair comparison?
The decisive dense comparison controls both stored capacity and the capability actually active for one task, then measures quality, throughput, memory and energy together.
For each question, we publish the test conditions, the observed result and the conclusion it supports.
What failed, and what we learned
Several earlier schedule comparisons were stopped or invalidated before they could support a trustworthy conclusion.
One comparator defect was repaired under a new experiment identity. The corrected comparison completed and provided no basis for adopting the proposed schedule change. Both records remain available.
Negative results guide the next experiment. We retain failures and keep changing training evaluations separate from fixed comparisons so another reviewer can follow the decision.
Two sponsored research pathways
Native ERAIS Model Research tests a fixed native model against an agreed baseline on a partner’s domain and compute budget.
Governed Conversion & Adaptive Serving Research investigates how much behaviour a supported conversion retains and how quality changes with selective execution.
Both can be scoped through a Research & Benchmark Partnership. Bring a model, task or research question, the comparison that matters to you and the resources available for testing it.
Inspect the work
Start with the ERAIS overview for the architecture and selected results.
Progress and proof dates each engineering review and research snapshot. The evidence library keeps methods, limitations and retained failures available for deeper review.
Additional implementation and reproduction material can be reviewed through technical diligence, subject to source licences and confidentiality.
Before publishing a performance claim
We define the question, baseline, data, quality threshold, environment and decision criteria before the comparison. The review checks parity, repeated measurements, exact source identity and adverse results.
Read the benchmark policy, method, roadmap and complete claim register for the protocols and decisions.