Experiment

An experiment is the atomic unit NeuronScope produces. Every artifact, every diff, every report, every research write-up is anchored to one or more experiments — never to a live process.

Anatomy of an experiment

• Inputs. A model, a dataset or prompt, a probe, and a captured environment. Each is fingerprinted before recording begins.
• Plan. A frozen description of which layers, tensors, and summaries will be observed. Fingerprinted separately from the model.
• Observations. The tensors and metadata produced during the forward pass.
• Artifact. The sealed .nsz that binds all of the above together.

Lifecycle

An experiment moves through five states: resolved, planned,instrumented, running, sealed. Each transition is checkpointed so a failure is attributable to a specific stage — see thearchitecture page for the flow diagram.

Constructing one

import neuronscope as ns

with ns.record(
    model=model,
    probe=ns.probes.LayerActivations("encoder.*"),
    out="run.nsz",
) as run:
    run.forward(**inputs)

print(run.artifact.fingerprint)

Sessions vs experiments

A Session groups related experiments — for example, one per epoch during training. Sessions do not add semantics; they are a folder-level container so batch analyses can iterate cleanly. If you have a single recording, you have a single experiment and can ignore sessions entirely.