Improve (workflow)

Stage 5 — Improve. Tune prompts, judges, and configurations. Re-run.

Stage 5. With a baseline score in hand, improve by changing what you control: prompts, retrieval, model, judge rubrics. Then re-run.

The question this stage answers

"Can I make it better — and prove it?"

What to do

1. Pick one variable. Don't tune everything at once.
2. Make the change. New prompt, swap model, tweak temperature, add a retrieval re-ranker.
3. Re-run the evaluation. Same scorers, same judges.
4. Compare to baseline. Stratix shows per-dimension deltas.
5. Iterate. Keep what wins; revert what doesn't.
6. Optimize the judges themselves. GEPA tunes the judge rubric against labels.

Two kinds of improvement

• System under test — the model, prompt, retrieval, etc.
• Measurement system — the scorers and judges themselves

Both deserve attention. Improving the measurement system (GEPA) often surfaces issues in the system under test you couldn't see before.

Common pitfalls

• Tuning to the eval, not the customer. If your eval doesn't reflect customer reality, you'll over-fit.
• Not retiring obsolete examples. Your eval set should evolve as the product evolves.
• Treating one improvement as final. The frontier moves; revisit quarterly.

Where to next

• Govern →
• Tutorial: Optimize a judge with GEPA
• Stratix Premium — Judge optimization
• Concept: Judges