Prompt-eval & VLM
VERIFIED Multimodal PromptSet evaluation produces Detection / TriageCase artifacts.
- Dual prompt registries exist (keep them consistent when changing prompts)
- Confidence gate 0.5
- VLM call timeout 60s
- Guardrails aligned with Promptdesk standard (no API keys in prompts; fail closed)
NOT FOUND Offline eval harness as a shipping product feature.