Metrics

Maturity: verified

Metrics are structured observations emitted during execution and associated with a defined run, attempt, and step.

Optimizer-step metrics

The application emits one strict JSON object for each completed optimizer step.

Field	Type	Meaning
event_type	string	always optimizer_step
global_step	positive integer	completed optimizer-update count
loss	finite number	masked next-patch-prediction mean-squared error
num_target_patches	non-negative integer	valid source/target patch pairs contributing to loss
sample_ids	array of strings	ordered samples consumed by this step

event_type, global_step, and loss are application-owned. The NPP component supplies num_target_patches; collisions with owned names fail. Metrics are serialized with non-finite values forbidden and atomically published at the attempt’s terminal state.

There is no validation loss, accuracy, confusion matrix, signal-to-noise-ratio curve, benchmark score, learning-rate series, throughput result, or distributed reduction in the current application.

See Also

• Logging and observability

• Experiment tracking