Requirements

Define a requirement so the author and reviewer agree on what the evidence must demonstrate. State the expected outcome and the conditions under which it must hold.

For example, Quantized classifier performance can require both acceptable accuracy and latency. Keep these together when they support one decision about the model, and give each measurement its own check.

Use Checks to define how the results are assessed and Submissions to follow the evidence and its review status. Open a submission for its saved revisions and Conversation history. A new model candidate can have its own submission against the same requirement.

Create a requirement#

Open Requirements in your project and select Add requirement.

Make the description specific enough for someone else to assess the evidence. Include the target, conditions, and acceptance criteria. For example:

The quantized classifier must achieve at least 90% accuracy on validation_v1 and a p95 inference latency of at most 20 ms on target_device_a at batch size 1.

Reference a requirement from another system#

Reference a requirement when your team already manages it elsewhere, so reviewers can trace the evidence back to its source.

In Add requirement, choose Add from another system and enter its source URL, name, and description.

Keep the name and description up to date when the source requirement changes; these fields are maintained manually in the Hub.

Manage requirements#

Open a requirement to edit its name and description. Use Owner to identify who is responsible for it and Tags to organize related requirements. Assigning an owner does not change who can view or edit the requirement; project permissions control access.

Next steps#

• Add checks to define how evidence will be assessed.
• Submit evidence from the relevant runs. One submission can include a profile run and an evaluation run for the same model.