Crate nexcore_cognition 

nexcore-cognition

Typed cognitive engine — the transformer algorithm as strict Rust.

Meta-cognitive origin

This crate captures the fundamental algorithm behind large language model cognition: attention selects, transformation processes, generation builds. Each module maps to an observable pattern in how neural networks process information, translated faithfully into Rust’s type system.

Architecture (bottom-up)

pipeline ──► generator ──► block ──► attention + feed_forward
                             │          │            │
                             ▼          ▼            ▼
                         normalize   mask         tensor
                         residual                   │
                         embedding                error

T1 Primitive grounding

Module	Primitives	Cognitive role
tensor	N, Σ, ×, ∂, κ	Numerical substrate
embedding	μ, λ, N	Symbol → vector
attention	κ, →, N, μ, Σ	Relevance selection
feed_forward	μ, ς	Nonlinear transformation
residual	π, Σ	Context preservation
normalize	∂, N	Signal stability
block	σ, ∃	Composable unit
mask	∂, →, ∝	Causal constraint
generator	σ, ρ, ∝, →, ∂	Autoregressive output
sample	N, ∂, ν, κ	Stochastic selection
metrics	κ, N, ν, μ	Self-measurement
pipeline	σ, →, Σ, κ	Full cognitive flow

Modules

attention

Scaled dot-product attention and multi-head attention.

block

Transformer block — the composable unit of cognition.

embedding

Token embedding and positional encoding.

error

Error types for the cognitive engine.

feed_forward

Feed-forward network (FFN).

generator

Autoregressive generation — building output token by token.

mask

Causal masking for autoregressive attention.

metrics

Meta-cognitive self-measurement.

normalize

Layer normalization.

pipeline

Full cognitive pipeline — the complete flow from input to output.

residual

Residual connections (skip connections).

sample

Sampling strategies for token generation.

tensor

Dense tensor type and operations.

Functions

make_rng

Create a seeded or OS-random StdRng for use with the cognitive engine.