Fracture / model conversion
Your model represents an investment in data, training and development. Fracture converts supported components into native ERAIS execution so that capability can become part of your next system.
Start with compatibility and the behaviour you need to retain. We agree the conversion scope, comparison and target hardware before the project begins.
Connections, weights, state and update rules.
Compare outputs and state, then measure execution cost.
Connect memory, adaptation and action, with tests for the combined system.
01 / Supported sources and research
Source structure, learned weights, recurrent state and update rules determine the conversion. Each record identifies what was loaded, preserved and tested.
Imported and running
Male CNS
The current male import contains 166,700 neuronal entries and 25,582,938 directed pairs. Native execution retains that declared graph. The fly and Doom experiments specify their own dynamics, sensory mappings and controls.
Read the declared fly modelSource record
Connectivity comes from the published source. Neuronal dynamics and actuator mappings are modelling choices, with their own tests and assumptions.
Imported and running
Female CNS
The female BANC import is loaded in a separate executor with 188,508 neuronal entries and 13,620,865 directed pairs. Its source identity and retention record are independent of the male model.
Read the current source recordSource record
The current check establishes the loaded source and reported retention. Behavioural validation and matched runtime comparisons remain specific to this model and its chosen dynamics.
Recorded conversion results
Trained AI models
The recorded Qwen3 0.6B conversion reorganised learned feed-forward components into experts and checked reconstruction with the complete expert bank. Supported speech and game-control conversions extend the work into other modalities.
Inspect the Qwen reconstruction recordSource record
This record applies to the named Qwen model, converted layers and engineering cases. Each additional model family is evaluated on its own tasks.
Qwen complete-expert reconstruction, with the tested layers and output comparisons published.
Inspect this path Component conversionRecorded LuxTTS/Vocos projection experiments and a separately qualified K7 listening route.
Inspect this path Recorded probesSaved image-classification examples and the conditions under which they were produced.
Inspect this path Embodied researchNative game-control components and live agents connect model outputs to visible actions.
Inspect this path02 / Recorded reconstruction
Real Qwen3 0.6B fracture, exact K32 engineering reconstruction, and a completed protected four-way trace for material-sparsity alignment. The figures below describe the recorded Qwen3 0.6B engineering cases.
Complete expert bank / K32
This reconstruction used all 32 experts. The diagram shows that bank’s recorded structure.
Full-compute reference.frozen engineering cases
matched exactly
| Signal | Matched | Compared |
|---|---|---|
| First-token logits | 5 | 5 |
| First-token top five | 5 | 5 |
| Greedy generations | 5 | 5 |
36033f8467ecThe bounded alignment path is executable, but no complete aligned checkpoint is promoted; K8/K16/K24 quality, savings, broader generalisation, and production readiness remain unevaluated.
03 / Integration
A preserved component becomes useful when it can contribute to a larger system. The cyborg experiment connects a complete native neuronal expert to an adaptive ERAIS executive.
The expert’s source weights are frozen and its fast state is episode-owned. Actual descending-neuron activity contributes to movement. The executive retains its own learned visual, navigation and associative state. Each part has a declared role and clock.
Explore the architectureConsumed image inputs are bound to the corresponding model state.
Executive associations and outcome memory persist across attempts.
Native neuronal counts and the executive contribute through an engineered motor interface.
A useful comparison identifies the source, inputs, initial state, hardware and workload. We assess quality alongside wall time, CPU time and memory. That is how an engineering result becomes a sound deployment decision.
04 / Live experiments
Open the reference, native network or cyborg view. Each ongoing run has its own history; the panels report source-owned activity and current gameplay.
01 / Reference dynamics
The retained MaleCNS network and original v6 learning rule.
Snapshots are off. Start viewing to connect.
Open to connect to this controller's observed activity.
02 / Compiled dynamics
The retained network and learning rule in the compiled ERAIS bank, with adaptive neural cadence.
Snapshots are off. Start viewing to connect.
Open to connect to this controller's observed activity.
03 / Adaptive controller
A native neuronal expert alongside ERAIS visual circuits, adaptive association and persistent memory. Source status reports the active composition.
Snapshots are off. Start viewing to connect.
Open to connect to this controller's observed activity.
Freshness-checked snapshots, at most four requests per second per visible player. Not continuous video. Pixels and telemetry come from the same acquisition; expired frames disappear. Pausing a view does not stop its player.
A dash means telemetry has not been received or that field was not reported. Player kills require engine-reported credit, not merely an enemy death. Round averages exclude interrupted attempts and show their own sample counts. RSS is resident memory; the additional footprint includes proportional resident and swapped pages. These are whole-process observations, not isolated kernel costs.
The views use the same campaign family but have independent progress and histories. Do not divide their current numbers to claim a conversion speedup or a learning advantage. Viewing controls cannot send game inputs.
ViZDoom and Freedoom assets. Not affiliated with id Software. The portrait illustrations are not biological measurements.
Your model and workload
Tell us which model you use, what it must do and where it needs to run. We’ll assess support and define a conversion project with clear deliverables and a direct source comparison.