Specification-to-execution
A fuzzy goal becomes a repeatable workflow or tool plan.
Risk: optimizing the wrong specification.
Mechanism
Five loops turn “self-improvement” into inspectable mechanisms.
Most projects fail at verification or retention. These five loops are review lenses for asking what changed, what selected it, what verified it, and how the result survived.
Search
The system proposes multiple candidates and selects among prompts, code, architectures, routing, or parameters.
Risk: random walk or benchmark overfitting.
Evaluator
Tests, benchmarks, hidden tasks, or judges decide whether a candidate improves the system.
Risk: Goodhart, leakage, or weak tests.
Reflection
Failed trajectories become typed memories, lessons, or skill notes that affect future runs.
Risk: memory pollution and false lessons.
Population/archive
Multiple variants survive across generations with lineage and diversity pressure.
Risk: premature collapse or diversity without useful selection.
Evidence and limits
The loops are lenses for reading systems.
The five-loop framing is a reader-facing compression of the survey methods, project cards, and benchmark discussion. It helps compare systems, but individual claims still need source links, evaluator details, and negative examples.