voice-evalslisted

# voice-evals — evals for phone voice agents Two jobs, both against the same set: **write cases** (§ Cases) and **score a round** (§ Scoring). Tell the user in one sentence which one is running. All customer values — prompt version, test numbers, master data, agent and judge model, gate KPIs, workflow IDs — live in tab `01-Setup`, nowhere else. The data available only determines where findings come from, never the procedure: runs from `04-Läufe` › pasted-in transcripts › audit checklist alone. The thinner the data, the more findings come from `references/rules.md` §1 instead of observation — flag that in the result. ## Why the rules exist The checklists live in `references/rules.md`, the scaffold in `references/template.md`. Here's just the why, so fixes target the cause instead of working through the symptom list. The § references point into `references/rules.md`. 1. **STT and TTS have separate failure modes → mirrored rules.** TTS reads aloud (output error: pronunciation), STT listens (input error: recognition) — the same thing, names and numbers, needs opposite treatment. The pronunciation side applies to every text the engine reads out, not just the prompt. §1.3 2. **Instruction density lowers compliance.** Every rule exactly once, details in the knowledge base instead of the prompt. 800 words without duplication beat 400 with. §1.4, §1.5 3. **The prompt describes behavior, not a case directory.** An eval set covers two dozen cases, the agent experiences thousands.