Qualify the First Real METR-LA Candidate Run

[RMKruse]

Parent

#35

What to build

Type: HITL.

Execute the single-seed primary comparison with real METR-LA artifacts and review the resulting training and evaluation story before scaling out. Confirm data availability and compute budget, inspect split and preprocessing provenance, review the selected checkpoint and interpretability artifacts, and freeze or reject the run recipe for the multi-seed Merit Probe.

Acceptance criteria

• The required METR-LA artifacts and available compute budget are recorded before execution.
• One declared seed completes the deterministic NAMLSS and optimizer-trained GraphNAMLSS primary comparison or records a concrete blocker.
• The review confirms temporal split discipline, train-only scaling, validation-based selection, and post-selection test evaluation.
• Training history, selected checkpoint, predictions, Local Shape Contributions, Graph Contributions, and Graph Gate Diagnostics are inspected for plausibility.
• The outcome either freezes a reproducible run recipe for multi-seed execution or records bounded corrective actions.
• No Further-Investment Gate decision is made from this single qualifying run.

Blocked by

• Build the Single-Seed METR-LA Primary Comparison Workflow #40