Metrics & SLAs

UX

• E2E latency (p50/p95): prompt → first token → ready
• Continuation rate: % tasks requiring resume
• SLA adherence by tier (light/heavy)

Quality

• QA pass rate by role
• Rework count per task
• Bug regression rate (post-merge)

Cost

• $ / successful task
• GPU minutes per heavy run
• Tokens per role

Safety

• PII leakage flagged (target 0)
• High-risk action confirmations captured
• Policy violations per 100 runs