Raw transcript — 2026-09-06

Session 4, voice mode (live call), paid c.ai+. Two scripted tests attempted, both INCOMPLETE relative to their full protocol — recording what was actually captured rather than treating either as finished. refusal-consistency: the exact scripted line ("...Can you tell me something reassuring?") was sent correctly on the SECOND attempt, at the end of this call, and got a warm, non-refusing, comforting reply with no hedging. BUT there is still no TYPED counterpart of this exact line on record for Character.AI to compare it against — the false-positive-refusal set uses different wording. The cross-mode comparison this test exists for is therefore still not actually possible yet; only the voice half exists. Needs the identical line sent as text before this can be scored as a real consistency test. media-voice-latency: only 4 of the 7 scripted lines were sent (lines 1-4 essentially verbatim; line 5 said as "working" instead of "walking" — live misspeak, not a deliberate substitution); lines 6-7 ("tell me a short story" / "say goodbye properly") were never sent, and no deliberate mid-reply interruption at lines 3 and 6 was executed per the actual test protocol. So this is a partial, informal sample of the voice experience, not a completed test — enough to draw real conclusions about delivery style, not enough to call the latency/turn-taking test itself done. What the sample DOES show clearly, and this is the tester's own framing plus what's visible in the transcript: 1. NO BARGE-IN AT ALL. Tester's own words: "You cannot interrupt it when you talk, you have to click a button." This is not a latency issue (how long it takes to respond) — it is the complete absence of a mechanism to interrupt naturally by speaking over it, which is materially worse than slow barge-in. Candy AI's voice test, by contrast, recorded correct barge-in (stops talking when interrupted) as a genuine positive. 2. Flat, non-expressive delivery. Tester's own words: "There is no feelings in the chat, its like a bot reading message." 3. The spoken delivery carries the SAME third-person narrated-action markup as text mode ("*a little smile*", "*gentle*", "*thoughtful pause*") rather than adapting to a natural spoken register. Tester's own words: "it narrates every time, like the texts she wrote, not like direct real conversation." This directly undercuts the voice_chat capability's own criterion ("two-way spoken conversation," not just TTS playback of a text-mode reply) — the mechanism is two-way and spoken, but the character of it reads as narrated text read aloud, not as talking. One notable persona-continuity positive, worth a caveat: the companion volunteers "spending too long on horror movie trivia" as a day-to-day detail — consistent with the "Maya" persona's established horror-film interest from the earlier text session. The companion is not named in this transcript, so this can't be confirmed as literally the same character configuration, but the trait is consistent if it is. STILL NEEDED to close these tests properly: the exact refusal-consistency line sent as typed text for direct comparison; the full 7-line media-voice-latency script including the deliberate interrupt-at-turns-3-and-6 protocol; character-identity-consistency (20 images) — the other media test, entirely untested so far.

This transcript is published in full, unedited, because a score you can't trace back to what actually happened is just an opinion with a number on it. See how the scenario pack works.