← All questions
EvaluationHard· LLMOps· Cost Optimization· Attribution

Your eval budget is cut. A teammate wants to drop the '...

The remove-the-evidence ablation is necessary because no attribution judge can distinguish between load-bearing and redundant citations using only the fixed output. Dropping this test creates a blind spot in the evaluation harness. Instead of removing it, you should subsample a rotating fraction of the held-out slice or cut other, less critical line items like macro-level dashboards that can run at a reduced cadence.

Code playground (Python)