Operations

Every operation combines named EvaluationTargets with one strict TOML plan. Validate a plan before simulation:

julia --project=. bin/brainlesslab.jl check PLAN.toml
julia -t auto --project=. bin/brainlesslab.jl run PLAN.toml --root records

Choose the operation

Question	Operation	Interpretation
What dynamics occur in one composition?	ProfilePlan	descriptive profile
How do declared parameter cells behave?	SweepPlan	development grid
Is a registered mechanism necessary?	AblationPlan	paired intervention
Which registered parameters perform better on training targets?	EvolutionPlan	selected candidate
How do declared conditions behave within tasks?	BenchmarkPlan	task-specific statistics and contrasts

Profiles run declared analyses over raw trials. Sweeps pair seeds across explicit cells. Ablations include an implicit baseline and validate capabilities before execution. Evolution selects on training targets before held-out evaluation.

Benchmark cases

A multi-condition benchmark case must declare a baseline. The operation then reports paired within-task contrasts.

An anchor-only case has exactly one condition and omits baseline. It reports the condition’s raw and normalised intervals with an empty contrasts table. Task anchors are reference points, not comparison conditions.

No benchmark operation creates a cross-task aggregate. A leaderboard or qualification policy is a separate choice from registry membership and record publication.

Plans and experiments

Files under plans/examples/ are small executable examples. Versioned scientific protocols belong under experiments/ as ExperimentSpec bundles.

Source: src/operations/, src/records/PlanIO.jl, bin/brainlesslab.jl.