The conceptual heart of NeuronScope. Each chapter explains what a primitive is, why it exists, how it works, when to use it, and what it costs. Chapters cross-link to their API pages, examples, tutorials, and ADRs.
Deterministic identity for models, datasets, prompts, and environments.
The captured runtime — OS, Python, wheels, CUDA, seeds.
How NeuronScope observes model structure without owning it.
Streaming, chunked, and mapped datasets with reproducible ordering.
Structured prompts as first-class, hashable objects.
Adapters for PyTorch, JAX, TensorFlow, and custom runtimes.
A recorded run with pinned inputs and a portable artifact.
A repeatable, comparable measurement over one or more experiments.
Overhead, memory, and layer-level timing while recording.
The exact contract NeuronScope makes and what it explicitly does not.
Rendered summaries produced from one or more artifacts.
Structural and semantic checks over inputs and recorded runs.
Diffing two experiments and interpreting the delta.
The .nsz format, its stability guarantees, and forward/backward compatibility.
How adapters, probes, backends, and plugins register themselves.
The fluent builder used to construct experiments and benchmarks.
The plugin contract, discovery, isolation, and versioning.