ce-optimize

Metric-driven iterative optimization loops are automated by running controlled experiments in isolated git worktrees and evaluating results against hard gates or LLM judge scores. The process converges on the best code or config variant by automatically comparing results against a baseline and tracking the best state.

An LLM judge evaluates semantic quality using rubric-based scoring and stratified sampling to select superior solutions. It acts as a measurement harness when a measurable objective requires human-like judgment rather than hard numerical metrics to assess output quality.

Running controlled experiments in isolation using git worktrees involves generating a hypothesis backlog and testing mutable files while keeping immutable files fixed. This safely manages experiment branches, logs results, and enables safe rollback to avoid regressions if a variant fails gating.

Yes, a clearly defined optimization goal and a repeatable measurement harness are required to use automated gating. The harness scores variants against hard gates or LLM judge criteria, enabling automatic baseline comparison, best-state tracking, and safe rollback to prevent regressions.

Tuning search relevance and clustering quality automatically requires defining a measurable objective and running iterative optimization loops. By applying stratified sampling and rubric-based scoring, the system evaluates variants against hard gates to converge on the superior solution.

You should not use metric-driven optimization loops when a project lacks a measurable objective, a repeatable measurement harness, or a clearly defined scope of mutable versus immutable files. Without these elements, the system cannot generate hypotheses, run controlled experiments, or gate results safely.