Training metrics

Every run directory is self-contained. The trainer writes JSON files that the audit and the exporter read; nothing depends on a tracking service.

File	Contents
metrics.jsonl	One event per line: train, validation, core, final_bpb, budget_anomaly
provenance.json, resume_<step>.json	The configuration, fingerprints, software and batch layout of each training segment
run_status.json, budget.json	Phase, final status and the cumulative compute ledger
val_<step>.json, core_<step>.json, core_final.json, bpb_final.json	Evaluation results at the scheduled points
checkpoint_<step>/, latest.json	Checkpoints every 5,000 updates and at the end

train events carry the optimizer step, input tokens, loss and auxiliary loss, gradient norm, learning-rate multiplier and the cumulative theoretical FLOPs with their deviation from the admitted proof; methods add their own diagnostics (sampled depth, router statistics, sparse selection). Training logs every 10 updates by default (--log-every), plus the first and the last.

Optional tracking

Weights & Biases logging is off by default (--wandb-mode disabled). offline and online modes log numbers, the scientific configuration and provenance from rank 0; models, data, source and console text are never uploaded. --wandb-metrics core uploads a compact profile and full every numerical field. Keys carry event prefixes such as train/loss, validation/val_bpb, core/core_metric and final_bpb/val_bpb. Credentials come from the usual WANDB_API_KEY or a mounted netrc file, never from configuration files.