Nightly eval regression detected (2026-05-02)

[github-actions]

Regression Report: math-20

Baseline: 0.2.0-cerebras-llama3.1-8b
Current: ci-nightly
Status: REGRESSION DETECTED

Metrics

Metric	Baseline	Current	Change
accuracy	0.9000	0.8500	-5.6%
avg_latency	11.2776	3.6134	-68.0%
total_tokens	3194.0000	1967.0000	-38.4%
avg_tool_accuracy	0.0000	0.0000	N/A

Alerts

• [WARNING] math-20/accuracy: 0.9000 -> 0.8500 (-5.6%, threshold: 5%)