Four-panel dashboard from synthetic evaluation data. v3 has the highest end-to-end success rate (170/200, 85.0%). v3 has the largest completed-run p95 latency (4600 ms) and a 5.0% timeout rate. The largest off-diagonal v3 routing confusion is true technical predicted as refund (8 tasks). Cost-success Pareto candidates are v1, v3. Success intervals are Wilson 95% confidence intervals; routing cells show counts and row percentages.