Make upgrades manageable
Integrate a new specialist, tool or memory and directly test the existing behaviour it affects.
ERAIS / Efficient Routed Adaptive Inference System
Build intelligence from reusable capabilities.
ERAIS brings specialist components into one architecture. Native language training starts with raw bytes; Fracture brings in skills learned by existing models. We’re testing how language, vision, audio, reasoning and memory can work together.
Connect the relevant expertise, carry state forward and account for the work performed.
The goal is growing capability within a practical operating budget.
See what has runWhy it matters
Make capable AI practical in more places.
Selective execution is a route to investigating lower operating costs and deploying useful capability across a wider range of hardware.
Bring AI to the work
Move capable systems from datacentres toward workstations, field equipment and edge devices.
Inform future hardware
Use measured access patterns and execution costs to identify what the architecture needs from the machine.
These goals connect the architecture to a commercial question: how much useful capability can a team develop and operate within its budget?
Get help with your AI projectModular development
Extend a system with a tested component.
Test an audio, vision or specialist addition as a separate candidate. Check when it should run and compare the existing language behaviour before and after the change.
A recorded integration experiment trained with real audio and labelled video while preserving the frozen language core and inactive capability exactly in the tested configuration. That gives us a concrete starting point for testing modular upgrades.
See how this was testedEngineering taskDevelopment mechanism
Add a capability
Integrate a candidate specialistPreserve required behaviour
Compare against the retained referenceUnderstand a regression
Trace the change, inputs and outputsRecover a working system
Restore its recorded state and configurationExecution evidence
Source models and measured engineering results.
The current review covers retained connectome models, exact-state checks and runtime improvements. Earlier training and conversion records remain available below.
Engineering evidence · checked 2026-09-20
Current systems and measured improvements.
Source models, execution checks and component timings from the latest engineering review.
MaleCNS source model
166,700 retained neurons
25,582,938 directed pairs
malecns-v1.0 is loaded in a running executor. The source import reports complete retention of its declared network.
Model dynamics, input mappings and anatomical coordinate coverage are specified separately.Female BANC source model
188,508 retained neurons
13,620,865 directed pairs
banc-888-v3 is loaded in a running executor. The source import reports complete retention of its declared network.
Model dynamics, input mappings and anatomical coordinate coverage are specified separately.Articulated-body execution
30.3% lower CPU time
935.1 → 651.8 ms per 20 ms simulated interval in the recorded laptop comparison.
Same 200 physics steps per interval. Eight measured intervals per arm; shared-host component timing.Native neuronal execution
6.2–7.1% lower CPU time
The narrow prefetch update improved both recorded laptop tapes. Complete mutable state and weights matched byte-for-byte.
Neural-kernel measurements with frozen and learning-enabled profiles. Whole-loop cost includes physics and other work.Methods, provenance and remaining work
One shared-laptop ABBA comparison, eight measured intervals per arm, with retained physics cadence. Component timing.
Fixed retained-image tapes, eight measured intervals per variant. Complete state and weights compared byte-for-byte over four intervals. Neural component timing.
Actual simulator wiring, source identity, complete activity, image hashes and clean checkpoint/exit were checked.
- Whole-CNS articulated simulation remains below real time on the measured laptop.
- Calibrated muscle and CNS proprioception mappings remain research work.
- Receptor-state images have incomplete spatial support.
- Model quality and end-to-end commercial savings require workload-specific comparisons.
ERAIS source: 6f20efa8bf6f6f2a86312d0fb80eef1460bb19a1. Each measurement retains its source receipt digest in the downloadable record.
Earlier training and conversion records
3.40×descriptive same-host throughput ratio
68.45% lessCPU-package energy per reported valid target
A same-host developmental diagnostic, not a matched-quality control, paper reproduction, whole-system energy measurement, or model-superiority result.
Fracture component: less execution overhead
Selected Gemma MLP execution measured 1.982× the speed of its previous implementation on 48 retained rows, with unchanged output for those rows.
Cloud CPU choice: a faster exact continuation
Two cloud CPU configurations resumed the same retained 3M checkpoint and completed the same 27-update interval. The faster configuration measured 1.390163× the training throughput.
Local CPU: a clear small-scale systems signal
At TinyStories paper-text scale on the same CPU host, the corrected ERAIS full-stack window measured 3.40× the dense reference throughput and 68.45% lower CPU-package energy per reported valid target.
Cloud CPU: sustained full-stack continuation
A cloud CPU continuation completed 16.79M valid targets with exact terminal state, sustained throughput, zero swap events, and high use of its declared thread budget.
Cloud GPU: full-stack execution at higher throughput
The repaired TinyStories full stack completed 205 updates on a T4 with bounded quality movement, active cognitive paths, compact sparse execution, and an exact-resume checkpoint.
Cloud GPU: materially more memory headroom
A governed T4 training intervention reduced peak accelerator memory by 45.13% while retaining more than 100,000 valid targets/s and terminal quality inside the registered window.
Stored scale: feasibility, headroom, and the next bottleneck
At 1.61B stored-parameter scale, a larger-batch A10G arm completed ten finite updates, exercised the full governed stack, retained exact restartability, and exposed GPU occupancy as the next optimisation target.
Native + Fracture + worldTrain from raw bytes. Reorganise learned capability into experts. Bring language, speech and vision into one adaptive system.
These completed tests give the programme a foundation. Integration and trained capability are qualified for each model profile.
Inspect the systems in operation.
The public demo connects measured Fracture, native-byte language, sparse vision and symbolic reasoning results without exposing the proprietary engine.
Explore the ERAIS demosNative training snapshot · published 4 September 2026
The next step is useful quality at larger scale.
This completed adaptive raw-byte run used 87.3M parameters with sparse K1–2 routing active. It records one trained profile. Longer matched experiments test where quality and active-compute advantages emerge; their results are assessed separately.
87.3M parametersnative adaptive byte architecture
2,048 updatescompleted training updates
3.4964 BPBterminal source measurement
K1–2 activesparse routing stayed operational
Does the learning transfer?
Measure improvement on held-out tasks and a fixed evaluation chosen before the result.
Can it compress safely?
Can routine input become less work without losing meaning, order or useful surprises?
Can it know more without doing more?
Can stored capability grow while the work performed for one task stays under control?
Does it survive a fair comparison?
Does any advantage remain against a conventional model under the same declared rules?
Quality and cost decide where the system is useful.
The commercial opportunity grows when a model can deliver the required result on a workstation, product or field device at an acceptable operating cost.
One architecture / many ways to work
Real work is read, heard, seen and done in motion.
ERAIS is being built to connect language, voice, vision, memory and action through shared state and explicit component interfaces.
Private work on a workstation
Read procedures, compare evidence and prepare a decision without sending every task to a giant datacentre.
Hands-busy field support
Hear a spoken note or an alarm, then bring in only the knowledge needed for that moment.
Visual inspection
See an asset or defect, compare the relevant evidence and leave unrelated parts of the system idle.
Execution and evidence tools
Follow a result back to the work that produced it.
The ERAIS Observatory keeps each result connected to what changed, what ran, what failed and what happened next.
That means less time reconstructing an experiment from scattered logs. A result can be explained, repeated, repaired or rejected from the record that produced it.
See how NeuroForge tests its own workResult traceConnected
- Capability
- Real audio + labelled video
- Core check
- Language core + other capability exact
- Evidence
- Metrics · artifacts · failures
- Decision
- Keep · repair · restore
01Isolate
Keep the accepted system still.
02Observe
Record what actually ran.
03Explain
Find the part that changed.
04Decide
Keep, repair or restore.
Fracture / the migration bridge
Keep what today's models have already learned.
Fracture brings learned capability into ERAIS execution, from converted speech and language components to complete supported game-control policies. It builds on knowledge already captured in existing models.
The 11 September runtime record describes native text attention and KV execution with full-K converted language MLPs, alongside the donor tokenizer, embeddings and multimodal wrapping. The API guide explains how discovery identifies the selected dense, hybrid Fracture or native profile and its limits. The recorded native cutover shares the World owner with the public API and Discord adapters; it does not establish signed Discord acceptance, assistant quality, sparse savings or paid readiness.
Story Lab uses a separately trained raw-byte language checkpoint. The Minecraft player has its own perception-and-action runtime.
Implementation, integration and trained capability are separate milestones. Each checkpoint uses a declared profile; reconstruction establishes an exactness baseline, and recovery and adaptive execution must earn their own quality and resource results.
Donorstart with capability already learned
Fracturereorganise it into modular expert structure
Recover + adaptrestore quality and continue native ERAIS learning
Scale nextcarry the result across complete models and modalities
Build the next cognitive architecture
Build the next stage with us.
We’re looking for research collaborators, model builders and investors who want to test the architecture on consequential tasks and develop its commercial applications.