NeuroBrix is built on three inviolable design invariants that guarantee predictable, debuggable, universal behavior.
No hardcoded values anywhere in the runtime.
Every value comes from configuration files:
topology.json — execution flow and component connectionscomponents/<name>/runtime.json — component runtime configurationcomponents/<name>/graph.json — tensor computation graphsdefaults.json — user-tunable parameters (steps, guidance, temperature)hardware/<profile>.yml — GPU configurationsInline constants like sample_size = 128 are forbidden. This means the same runtime code handles image generation, LLMs, audio, and video models without any model-specific logic.
Missing data = explicit crash, never silent defaults.
If a required configuration value is missing, NeuroBrix crashes immediately with a clear error message explaining exactly what's missing and where it should be defined.
"Explicit errors allow for fast debugging. A default value can mask a bug for weeks."
This is intentional: silent fallback values are the source of hard-to-debug issues in production. A model that "mostly works" is worse than a model that clearly fails with actionable errors.
The runtime knows tensors, not domains.
NeuroBrix's runtime has no concept of "image", "text", "audio", or any domain. It operates on abstract tensors:
Tensor [B, C, H, W]Tensor [B, SeqLen]floatAll domain-specific behavior is encoded in the .nbx container's topology and component graphs. This is what makes NeuroBrix truly universal — the same runtime handles any modality.
These three principles work together:
The result: universal execution across model families without modification, precise error messaging, predictable behavior without hidden code paths, and zero runtime changes when adding new models.
Next: Model Hub