← Back to blog

EVOLVE — a self-improvement layer that's also a tripwire

How we let prompts improve themselves without losing oversight, money, or auditability.

Reginald Chan · 06 May 2026

EVOLVE runs GEPA experiments on Class-A modules, never on Class-B or Class-C. Every graduation requires an owner decision. Every margin must clear 60%. Every experiment is logged with a baseline_score, candidate_score, p_value, and a reflection rationale. Inconclusive results stay in the inbox until you choose. Nothing ships silently.