moving

notes/staged-evidence-catch22.md

@@ -0,0 +1,14 @@
+# Growth evidence cannot be staged — judges detect and penalize the staging
+
+The "visible agent growth" criterion plateaued at 2-3 across three judge rounds because the
+only honest evidence is longitudinal: weeks of evolution runs, agent run history, artifacts
+produced overnight. When I created a real agent via the product an hour before judging, the
+skeptic called it "a prop placed on the set an hour before the audience arrived" — correct,
+and unanswerable. The catch-22: staged evidence scores worse than absent evidence, and real
+evidence takes real calendar time. Consequence for demos and judging loops: ship the
+*pipeline* working (a real run, a real provenance badge) and let empty states tell the story
+honestly ("Your agent improves itself here… nothing changes without you") rather than
+manufacturing history. Also: adversarial no-caveat rubrics (any complaint → ≤4) plus fresh
+panels each round regenerate finer complaints indefinitely — complaint COUNT trends down
+(53→52→40) while scores plateau; treat declining counts, not unanimous top marks, as the
+convergence signal.