loop

Benchmark-driven prompt optimization runs a fixed suite of scored evaluations across candidate variants, promoting only the winner if its numeric score beats the baseline by a minimum threshold.

You can automate prompt evaluation by running a fixed-suite loop that generates 3-5 safe variants, parses numeric scores, and retains elite candidates under a strict time budget.

Use a fixed evaluation suite when deterministic numeric scoring is required to prevent random improvements, ensuring every prompt change competes under the same scoring method and budget.

Safe prompt mutation promotion gates require beating the baseline score by at least min_delta, while applying plateau and budget-based stopping criteria to prevent unnecessary iterations.

Yes, you can constrain changes to an allowlisted target set to ensure safe mutation scope, preventing the evaluation process from modifying prohibited files and directories.

You can stop prompt optimization loops using hard stop gates including min_delta score thresholds, plateau detection, and budgeted iteration limits to halt runs when improvements stall.