Choose an operation

An operation records why named conditions are being evaluated. The active runner uses one Plan representation for profiles, sweeps, ablations, benchmarks and calibration. Its operation field does not create a different neural update rule.

Choose the question

Operation	Question	Main output
profile	How does one composition behave across the declared protocol?	Native trial outcomes and block summaries
sweep	How do explicit parameter cells differ on development worlds?	Named cell summaries
ablate	What changes when a declared mechanism is disabled?	Baseline, intervention summaries and paired raw contrasts
benchmark	How does a model perform on the four task coordinates?	Task-specific reservoir and reference-null results
calibrate	What does the declared null policy score under this task contract?	Frozen task calibration

profile, sweep, ablate and benchmark are constructors that return a plan. execute runs a plan with an execution policy. The convenience function calibrate executes its fixed calibration protocol and returns an EvaluationResult.

The benchmark API measures task performance. It is not a speed test. Use runtime performance for compilation, throughput and recording costs.

Resolve a development sweep

from brainlesslab import (
    CompositionSpec, EvaluationSpec, NodeSpec, TaskSpec, resolve, sweep,
)

plan = sweep(
    CompositionSpec(NodeSpec("falandays"), TaskSpec("delayed_cue"), count=32),
    {"leak": (0.1, 0.25, 0.5)},
    evaluation=EvaluationSpec(blocks=2, trials_per_block=4, root_seed=85),
)
print([(target.target.id, target.model_config.leak) for target in resolve(plan).targets])

The Cartesian product of the explicit axes becomes named targets. The example contains three cells. It shares declared world streams across cells; different construction parameters may still produce different realised wiring. A selected cell is a development result, not a confirmed optimum.

The checked TOML example uses the same mechanism:

uv run brainlesslab check python/plans/sweep-delayed-cue.toml
uv run brainlesslab run python/plans/sweep-delayed-cue.toml --root records

Read the printed record, then inspect its native outcomes and seed ledger before choosing a follow-up. Selection and later confirmation must use disjoint seed partitions.

Treat controls as different mechanisms

The ablation helper supplies an implicit baseline and named intervention targets. It supports weight freeze, full-plasticity freeze, blind input and shuffled input at a declared absolute trial tick. Validation rejects redundant changes.

The benchmark helper supplies reservoir and reference-null targets for Tracking, Pong, CartPole and delayed cue. It preserves each task’s native units and default horizon. It does not average these coordinates into a competence score.

Empirical physical calibration uses exactly 1,024 independent reference-null trajectories in the calibration partition. Binary tasks declare analytic chance separately. Calibration is not parameter search or evidence of a model learning.

decode is an offline diagnostic over recorded events and trial metadata. It does not control the task or become an operation-specific plan format.

Evolution, general autodiff, durable checkpoint/resume, broad analysis registries and distributed execution are not implemented. Historical operations retain their original source revision. Continue with plan syntax, comparison controls and limits.