# Training loop

A visible zero-grad, forward, loss, backward, and optimizer-step sequence.

## Transcript

- **00:00:00.000–00:00:05.400** Zero gradients before the next optimization step.
- **00:00:05.400–00:00:10.800** The forward pass turns a batch into predictions.
- **00:00:10.800–00:00:16.200** Loss measures the gap between predictions and labels.
- **00:00:16.200–00:00:21.600** Backward computes gradients for each trainable parameter.
- **00:00:21.600–00:00:27.000** Step updates parameters; a lower held-out loss is evidence, not a guarantee.

## Accessibility and claim boundary

The poster is the complete static equivalent. Captions and transcript derive from the same reviewed Caption JSON source. Synthetic examples teach relationships; they do not certify a production detector.