Three-Layer Frustration Detection System
- Entity ID:
ent-20260419-238ef4de67e1 - Type:
concept - Scope:
shared - Status:
active - Aliases: three-layer frustration stack, frustration detection stack, Agnost AI frustration deep-dive
Description
Production instrumentation stack for detecting user dissatisfaction: (1) zero-latency regex classification on every user message for correction-tone phrases, (2) end-of-session LLM-labeled satisfaction (good/neutral/frustrated) trended across model versions, and (3) 0.5% survey sampling with full transcript upload on 'bad' rating. Extends the simpler regex-only view in prior rounds documented by the Frustration Telemetry entity.
Key claims
- 0.5% survey rate is a deliberate signal-vs-annoyance tradeoff
- Bad-rating surveys trigger automatic transcript upload
- ESC key count is tracked as an independent misalignment signal
Relations
- Three-Layer Frustration Detection System --[supersedes]--> Frustration Telemetry
- Agnost AI --[summarizes]--> Three-Layer Frustration Detection System
- Three-Layer Frustration Detection System --[supersedes]--> Frustration Telemetry
- userPromptKeywords.ts --[implements]--> Three-Layer Frustration Detection System