ADR-0003: Regression uses simulation, not live voice
Status: Accepted
Superseded by: —
Context
Section titled “Context”The five seed scenarios need to be re-run after a config change. Live voice is slow, expensive, and non-deterministic in ways that make failures hard to attribute.
Decision
Section titled “Decision”Do not use live voice for regression. The PRD specifies the ElevenLabs Conversation Simulation API against real middleware endpoints. This repo also runs HTTP middleware scenario tests in CI without ElevenLabs, because simulation is skippable when keys are absent. Simulation remains the intended full-agent check when configured.
Consequences
Section titled “Consequences”CI can prove gates, evidence rejection, and apply-time enforcement without a microphone. Full conversational regression still needs ELEVENLABS_API_KEY and an agent id. Known-good diffing of simulation JSON is coarse; named middleware assertions are the reliable bar.