Validation

Per-layer scientific-validation reports for the EMMA design.

Warning

Pre-implementation. Reports validate the proposed contracts and pinned rfgen recipes, not running code. The empirical tests they specify execute once src/emma/ lands, via the ported scientific-validation-orchestrator skill.

Who should read this. Reviewers who need to judge whether EMMA’s tasks, metrics, protocols, and datasets are scientifically defensible before implementation.

Before you start. Read methodology for the six-lens framework and the pre-implementation posture.

What this section is not. It is not a substitute for the reference contracts. It validates them; it does not restate them.

Per-layer reports

Reports mirror the build-plan layers. Load-bearing layers carry full six-lens reports; engineering layers carry all seven sections but concentrate depth on the construct, literature, and methodology lenses.

Layer	Report	Scope
primitives	primitives	enums, errors, the version and maturity model
schemas	schemas	serialized contracts (task, dataset, submission, run)
datasets	datasets	rfgen recipes, loaders, real-capture adapters, label extraction
metrics	metrics	per-task metrics, OOD-avg aggregation, sim-to-real gap
tasks	tasks	task construct validity and domain utility, OOD axes
harness	harness	the evaluator and the OOD protocol runner
leaderboard	leaderboard	held-out scoring, submission, anti-overfit
cli	cli	command surface and contract alignment
governance	governance	pre-registration, re-scoring, the ring-fence