Build a targeted defect test set and require: Which evaluation change is most defensible?

Rare severe conditions need targeted representative evidence and a metric tied to detecting the consequential failure, not aggregate accuracy.

The question

Preparing for AIGP? Take the free 5-min readiness quiz →

1. Increase the training set while retaining accuracy as the primary metric.

   More training data may help learning, but accuracy still obscures whether rare dangerous defects are detected reliably.
2. Replace real defects with synthetic defects and evaluate aggregate performance.

   Synthetic defects may support exploration but cannot alone establish real-world detection performance or realistic operating conditions.
3. Use random testing and report accuracy with confidence intervals.

   Random testing may contain too few failures to estimate detection performance, while accuracy remains dominated by nondefective units.
4. Build a targeted defect test set and require a detection-rate criterion. ✓

   Targeted testing supplies evidence for rare conditions, and a detection criterion reflects the severe consequence of missed defects.

How to remember it

Rare severe conditions need targeted representative evidence and a metric tied to detecting the consequential failure, not aggregate accuracy.

More Understanding How to Govern AI Development questions

• Version the model: Which documentation choice is most useful? →
• Provenance documentation alone establishes suitability: Considering provenance documentation specifically, →
• The validation population represents the new users: Which prior assumption requires reassessment first? →
• All 426 Understanding How to Govern AI Development questions →