Build the Single-Seed METR-LA Primary Comparison Workflow

[RMKruse]

Parent

#35

What to build

Type: AFK.

Build a config-driven single-seed METR-LA workflow for the Merit Probe Primary Comparison. It must run deterministic NAMLSS and optimizer-trained Tracer-Bullet GraphNAMLSS under one temporal split, feature, scaling, missingness, horizon, and metric protocol, then persist independently reviewable artifacts for both candidates.

Acceptance criteria

• The workflow consumes the existing METR-LA artifact contract and fails clearly when required local artifacts are absent.
• Deterministic NAMLSS and optimizer-trained GraphNAMLSS use the same temporal split, feature policy, scaling policy, missingness policy, forecast horizon, and metric definitions.
• Scaling and other learned preprocessing statistics are fit on the training partition only.
• The selected GraphNAMLSS checkpoint is fixed before test metrics are computed.
• Both candidates emit resolved configuration, provenance, metrics, predictions, and applicable interpretability artifacts.
• A small local fixture exercises the complete workflow in automated tests without downloading METR-LA or requiring benchmark-scale compute.

Blocked by

• Reload the Selected Checkpoint and Emit Interpretability Artifacts #39