Train an Optimizer-Trained GraphNAMLSS Synthetic Tracer Bullet

[RMKruse]

Parent

#35

What to build

Type: AFK.

Deliver the first complete optimizer-trained GraphNAMLSS path on a small synthetic traffic fixture. A config-driven run must update the local additive components, shared graph representation, parameter-specific graph heads, and Graph Contribution Gates through Gaussian location-scale NLL, select a candidate with validation evidence, and evaluate test data only after selection.

Acceptance criteria

• The named optimizer_trained_graph_namlss_training path is selectable through the experiment workflow.
• Optimization updates local additive components, the shared graph representation, both parameter heads, and Graph Contribution Gates from the declared objective.
• Every training and inference prediction satisfies the positive scale contract.
• Only training observations update parameters; validation evidence selects or stops the candidate; test evaluation occurs after selection.
• On a learnable synthetic fixture, the selected candidate improves the training objective over initialization without relying on a pinned optimization trajectory.
• The run emits resolved settings, training history, split metrics, predictions, and a selected checkpoint reference.
• Automated tests cover parameter updates, finite NLL, split discipline, and the end-to-end synthetic run.

Blocked by

• Review and Freeze the Concept Demonstrator #36
• Add a Multi-Seed Concept Demo Robustness Probe #37