On this page
The thesis
ERAIS is an expandable cognitive architecture. Language modelling is one thing it can do.
The question is simple to ask and difficult to prove: can an AI system keep adding knowledge, senses, tools and memory without making every task use everything it has?
Imagine you want to add one house to a neighbourhood. The brute-force approach would demolish the whole neighbourhood and rebuild it around the new house.
ERAIS implements a different approach: build the new house, connect it properly and leave the rest standing.
For each task, use the appropriate tools and effort. Adaptive routing and modular experts make this concrete. The research programme measures how quality and active computation change as the system grows.
What has run
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.
Architecture implementation, integration and trained capability are separate milestones. A supported path in the engine does not mean every checkpoint has been trained through it. The programme qualifies complete model conversion, sparse quality recovery, temporal reuse, capability-preserving growth and continuous multimodal operation on declared profiles, then records what each result earns.
The four gates that compound into a product
Does the learning hold beyond the training view?
Can the richer cognitive system learn something that holds reliably outside the changing view used during training?
Can routine detail be compressed without losing meaning?
Can frequent detail become fewer cognitive units without losing the order or novelty that cognition needs?
Can stored capability grow without making each task more expensive?
This is the central ERAIS scaling question. Does adding stored capability produce useful gains without proportionally increasing the work used for one task?
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.
The experiment behind each question remains controlled. The public programme explains the uncertainty, what ran and what conclusion the result earned.
Invalid work is stopped—not repackaged
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 rather than quietly changing the run. The corrected comparison completed, did not promote the proposed schedule change and narrowed the next question.
That behaviour is part of the product.
NeuroForge records what ran, keeps changing and fixed evaluations separate, preserves negative results and refuses to turn a broken comparison into a marketing claim.
Two sponsored research pathways
Native ERAIS Model Research tests whether one frozen native profile can clear a bounded domain gate against a qualified baseline, using a named compute and rights envelope.
Governed Conversion & Adaptive Serving Research investigates Fracture as a bridge from supported dense capability into ERAIS-native sparse structures.
Both normally operate through a qualified Research & Benchmark Partnership, connecting the strongest candidate directly to a partner workload and deployment boundary.
Proof without disclosure leakage
The public ERAIS overview carries the strongest bounded outcomes.
Progress and proof separates current systems receipts from the historical claim snapshot. The evidence library keeps methods, limitations and retained failures available for deeper review.
Implementation-sensitive routing, thresholds, model geometry, experiment recipes, checkpoint topology and protected infrastructure stay outside the unrestricted public site.
Controlled diligence is available when there is a legitimate review need and an appropriate boundary.
Promotion rule
Before a result can support a public claim, the programme freezes the question, baseline, data boundary, evaluation, quality threshold, environment, endpoint and all-outcome action tree.
Scout runs guide engineering. They do not become public performance claims without parity, repetition, source closure and review.
Use the benchmark policy, method, roadmap, and complete claim register for the current boundaries.