Benchmark an evolved CTRNN

Use this path after a SepCMA continuous-time recurrent neural network (CTRNN) record has completed. The benchmark names the saved model explicitly and evaluates it under a later protocol.

Do not copy coordinates into a new plan. A portable model reference retains the recorded model identity and its record location.

Create the model reference

SepCMA writes one model role named selected:

using BrainlessLab

evolution_record = "records/RECORD_ID"
model = BrainlessLab.Evolution.model_reference(evolution_record, "selected")

NSGA-II and CMA-ME do not create an implicit champion. For those strategies, choose an explicit ordered model ID, such as pareto-0001 or cell-0001, after inspecting the development record.

Build a frozen evaluation target

Use the same registered node design as the saved model:

composition = CompositionSpec(
    :structured_ctrnn_wall_benchmark,
    :compartmental_structured,
    :wall;
    n_nodes=60,
)

evaluation = EvaluationSpec(
    blocks=20,
    trials_per_block=1,
    horizon=1000,
    warmup=200,
    construction_scope=:trial,
    reset=:full,
    root_seed=0x3301,
    aggregate=:mean,
)

condition = EvaluationTarget(
    :saved_structured_ctrnn,
    composition,
    evaluation;
    model=model,
)

case = BrainlessLab.BenchmarkCasePlan(:wall, (condition,))
plan = BenchmarkPlan(:saved_structured_ctrnn_benchmark, (case,))

write_plan("plans/saved_structured_ctrnn_benchmark.toml", plan)

The example has one condition, so it reports condition statistics without a contrast. Add a declared baseline condition when the scientific question requires a comparison. Conditions in one benchmark case must use the same block structure and paired root seed.

The saved model was developed on Wall in the preceding tutorial. This protocol scores 800 ticks after warm-up, which exceeds Wall’s 200-scored-tick minimum. Wall remains a floor-check; the benchmark demonstrates portable model evaluation rather than broad competence.

Use evaluation seeds that were not used for search, selection, or a variance pilot.

Validate and run

julia --project=. bin/brainlesslab.jl check \
  plans/saved_structured_ctrnn_benchmark.toml

julia -t auto --project=. bin/brainlesslab.jl run \
  plans/saved_structured_ctrnn_benchmark.toml --root records

A successful benchmark record contains DONE, authoritative trial tables, summary statistics, the portable model reference, realised seeds, provenance, and checksums.

Interpret the record

The benchmark measures the named model under the declared task and evaluation protocol. It does not establish optimality, biological fidelity, cognition, or performance on another task.

One benchmark run is not automatically confirmed evidence. Use an ExperimentSpec, an appropriate control, independent randomised blocks, and the evidence ladder when the work supports a scientific claim.

Source: src/evolution/Evolution.jl, src/operations/Benchmark.jl, src/records/Records.jl.