plan-ml-experiment

name	plan-ml-experiment
description	Plan reproducible machine-learning experiments with leakage-safe random, grouped, or chronological splits; deterministic configuration fingerprints; dataset checksums; seeds, baselines, ablations, and artifact manifests.
origin	openai4s
category	reproducibility
capabilities	{"network":{"mode":"none","domains":[]}}

Plan an ML experiment

Use this skill before training begins. A reproducible experiment is a falsifiable question plus immutable inputs, a leakage-safe evaluation boundary, a declared metric, and enough recorded state to rerun the comparison.

Planning sequence

1. Write the hypothesis, intervention, baseline, primary metric, and decision rule before inspecting test performance.
2. Choose the unit that must remain independent. Use a grouped split for patients, molecules/scaffolds, sites, documents, or repeated measures; use a chronological split when deployment predicts the future.
3. Reserve the test set for the final comparison. Fit preprocessing and choose hyperparameters using training/validation data only.
4. Fix seeds, environments, input versions, and the exact configuration.
5. Run the baseline first, then one-factor ablations and the planned model.
6. Save per-example predictions and a manifest before drawing conclusions.

Deterministic helpers

from importlib import import_module

plan = import_module("plan-ml-experiment.kernel")
splits = plan.grouped_split(patient_ids, seed=42)
manifest = plan.experiment_manifest(
    config,
    data_paths=["data/cohort.csv"],
    seeds=[42, 43, 44],
    code_revision="<git commit>",
)

Use random_split(size, ...) only when rows are genuinely independent. grouped_split(groups, ...) keeps each group in exactly one partition. chronological_split(timestamps, ...) orders observations without shuffling. All return original row indices under train, validation, and test.

Minimum artifact set

• frozen config and config_fingerprint;
• source artifact/version plus SHA-256 checksums;
• code revision, runtime/environment versions, and random seeds;
• split indices or stable sample IDs and the grouping/time policy;
• baseline, ablation, and final per-example predictions;
• aggregate metrics with uncertainty and a failure analysis.

Determinism does not prove validity. Hardware kernels may remain nondeterministic, and repeating one biased split does not repair leakage or dataset shift. Report deviations from the plan rather than overwriting it.