Skip to content
BrainlessLab BrainlessLab BrainlessLab

Inspect and resume evolution

Use this path after an evolution operation stops before its declared iteration budget is complete. BrainlessLab can continue only from a complete generation boundary.

Inspect the record first

Use the directory printed by the original run command:

using BrainlessLab
using TOML


record_directory = "records/RECORD_ID"
manifest = TOML.parsefile(joinpath(record_directory, "record.toml"))


manifest["format"]
manifest["artifacts"]

Read request.toml, resolved.toml, and seeds.csv before continuation. These files show the requested plan, the resolved run configuration, and the realised random streams.

The status markers have precise meanings:

  • DONE means that the declared run completed;
  • INCOMPLETE means that a generation-boundary continuation point is available;
  • FAILED means that execution or record generation failed.

Do not edit the record or its checkpoint files. Resume validates the existing record before it writes more results.

Continue the same record

Call the public resume interface:

using BrainlessLab


record_directory = "records/RECORD_ID"
continued = BrainlessLab.Evolution.resume(
    record_directory;
    registry=DEFAULT_REGISTRY,
)


@assert continued.directory == record_directory

Evolution.resume continues the same directory to the original Evolution.RunConfig.iterations. It restores the last complete generation. Work from an incomplete generation runs again.

The call returns the same (result=result, directory=directory) shape as run_operation.

Check completion

After a successful continuation, the record contains DONE. It contains neither INCOMPLETE nor FAILED.

Check record.toml again because its artifact inventory and checksums now describe the completed record. Use the authoritative operation tables to inspect generation progress and recorded models.

Resuming a run preserves its declared protocol. It does not add an independent replicate, increase the evidence state, or validate the selected model.

Next, benchmark the saved SepCMA model.

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