evolve_closed_loop

Overview ¶

evolve_closed_loop exercises all 9 evolve interfaces in one runnable pipeline, using only the file-backed reference implementations plus a mock LLM client. Zero external deps, zero API keys, deterministic-ish (wall-clock used for Feedback.Timestamp; the pipeline logic itself is deterministic once inputs are fixed).

Scenario: carrier risk penalty evolution.

• Historical decisions were logged at a baseline penalty value.
• KPI feedback (on_time_rate) flows in.
• We ask an LLM for candidate penalty adjustments.
• WeightedEvaluator scores each candidate on a simple fitness model.
• ProposerFunc turns feedback into a proposal tied to the winning candidate.
• ShadowRunner compares baseline vs proposed fitness over the historical decision set.
• An approval gate (shadow delta + confidence) decides whether to Apply.

Run:

cd core && go run ./examples/evolve_closed_loop/