Persona Builder · Voice Mode

Which model should drive the builder?

Many LLMs through the same ten Sagar voice scenarios on the real CVI pipeline. The first run ranked four on checks, tool-calling, and time-to-first-token — and gpt-5.2 won cleanly. Then we shipped it, and it stalled 20–30 seconds mid-draft, went silent while writing the prompt, and stopped asking questions. None of that was in the chart. This run widens the field — the two Together reasoners re-served on Fireworks, plus frontier contenders from Anthropic, OpenAI, Google, GCP Vertex, Crusoe, and Cerebras — and adds the metrics that would have caught it.

New Charlie timeline rows are the only apples-to-apples comparison against the current builder prompt/dispatch architecture on sam/pal-maker-gemini: Grok 4.5, GLM-4.7, Gemini 3.5 Flash, and Gemma 4 31B. Older rows were recorded on the earlier prompt, so comparison against them is directional.
Standings clean = every check passed
Latency the TTFT chart hid median & worst · lower is better

Time-to-first-token measures when the model starts. It says nothing about when it finishes — and the builder's real cost is the long tail: the turn that draws the whole system prompt across a fan of tool calls. That turn runs silently while the creator waits. The worst-draft bar is the longest single drafting turn each model produced (full reply + all its tool calls), with the tool count on that turn.

Behavior the checks hid the model's own turns · higher is better
Scenario matrix
all checks clean partial pass never called tools dropped a turn ·not run yet