Module Responsibilities

Jsonnet (configs/) — structure and composition only. Every preset under configs/ablations/*.jsonnet produces a raw merged dict and computes its own run_dir. No validation, no types.

render (graphids/config/jsonnet.py) — _jsonnet C-binding call with typed tla_codes args (JSON-encoded so ints/bools/null round-trip correctly). Returns the rendered dict.

Pydantic / validate_config (graphids/config/schemas.py) — validation gate immediately after render. Catches null list fields, monitor/mode mismatches, un-namespaced class_paths, and LearningRateMonitor without a logger. Fails fast before any torch import.

build_run (graphids/orchestrate/instantiate.py) — imports class_paths via importlib, applies filter_kwargs against each target's __init__ signature, builds forced callbacks (ModelCheckpoint, EarlyStopping, MLflowTrainingCallback, VRAMDriftCallback when CUDA is available), and returns an InstantiatedRun(trainer, model, datamodule).

Stage primitives (graphids/orchestrate/stage.py) — build, train, evaluate. fit / test call these directly. No pipeline driver. Multi-stage chains are bash loops submitting each preset with SBATCH_DEP=afterok:<jid>.

SLURM (graphids/slurm/, scripts/run) — resource allocation and job submission. One script: scripts/run <preset.jsonnet> for training; scripts/run --mode {gpu|cpu} --command "..." for everything else. Reads configs/resources/submit_profiles.json (two entries only: gpu, cpu). Optional graphids.slurm.sizing.estimate_walltime_minutes queries MLflow history when invoked with --time-from-history.

The pipeline is strictly one-directional:

jsonnet renders (render)
    ↓
Pydantic validates (validate_config → ValidatedConfig)
    ↓
ResolvedConfig.from_rendered → build_run → (trainer, model, datamodule)
    ↓
trainer.fit / trainer.test

Authoritative detail: .claude/rules/config-system.md