Self-learning trend

Every reviewer edit and rejection feeds the skill-learning loop (Gate #3). These series should fall as promoted skill versions compound.

Reviewer rewriting (edit distance) fell 65% across 3 measured runs — the skill loop is compounding.

Feature rejection rate fell 74% across 5 measured runs — the skill loop is compounding.

Reviewer rewriting per run

How much reviewers rewrote generated content, run by run — this should fall as promoted skills compound. Select a run to inspect it.