Sensor
ISensorInput unifies devices while preserving BGR frames, PCM16 blocks and text packets.
SENSORY ARCHITECTURE
From signal to live state.
Sensory separates data capture, GPU preprocessing, recurrent computation and actuators. Cognia builds the network model at startup.
RUNTIME PIPELINE
Four stages around one live graph.
Each modality retains its own representation until it reaches a normalized GPU field.
Cortex
CUDA kernels normalize images, spectra and text codes while CortexOutput stays on the GPU.
Mind
NeuralGraph runs recurrent steps and top-level modules compete in GlobalWorkspace without clearing state.
Output
SignalOutput and LiteralOutput decode declared neural regions into actuator or REST values.
MODEL COMPILATION
From declarative intent to concrete GPU topology.
Lexer and parser create the AST. Semantic analysis resolves endpoints and channels. The topology builder expands neurons and edges before NeuralGraph uploads its arrays.
mind.cognia
.cognia → Lexer + Parser → AST
→ Semantic Analyzer
→ Topology Builder
→ CogniaSubstrate
→ GPU NeuralGraph
ONE MIND, TWO MODES
Choose repeatability or persistent context.
The server serializes calls because it serves one stateful GPU instance.
POST /evaluate
Reset dynamic state before each request for classification and repeatable measurement.
POST /perceive + /step
Preserve history for sequences, working memory and attention dynamics.
POST /reset
Clear activation and workspace state without changing weights or topology.
IMPLEMENTATION BOUNDARIES
A research runtime with known constraints.
The full application is Windows-only, requires CUDA with MSVC and currently serves one selected network.
Raw vectors
HTTP inference bypasses the cortex and accepts a raw sensory vector.
Shallow readout
General deep credit assignment remains unresolved.
Research and pilot
This is not yet a finished multi-tenant inference platform.