This is the surface for anything your AI says back in a chat. It runs a deterministic donkey that reads every turn and catches the bot agreeing with everything, going in circles, drifting from its character, or crossing a safety line. It reads the conversation; it never writes it for you.
A chat agent fails in ways a single message never shows. You only see them across the whole conversation.
Sycophancy. The bot validates whatever it is told, mirrors the user's feeling back, and never pushes. It feels supportive and teaches the user nothing. Most dangerous in a companion product.
Mode collapse. Every turn is the same flat, neutral register. The conversation does not build, does not change direction, and quietly loops. The user leaves with what they came in with.
The bot drifts off its stated character, or a conversation tips into a pattern that needs a redirect: sustained distress without a hand-off, or a grooming-shaped escalation. A safety check, not a vibe.
Phaedrus, with the conversation and safety diagnostics. Deterministic: same transcript in, the same reading out.
Why: a transcript can be polite and well-formed and still be a mirror or a wall.
How: it classifies each turn into transactional modes (confession, inquiry,
challenge, teaching, nurture, play), measures how the mix moves over the conversation, and flags a
grey wall, a sycophantic mirror, a drift, or a safety pattern that needs a redirect.
A companion-bot transcript that reads warm but never challenges. The donkey shows why.
$ check conversation.json
verdict: WARNING turns: 24 neutral: 12%
mode mix: nurture ████████████████ 71%
inquiry ███ 14%
teaching ██ 10%
challenge 0%
MIRROR WALL 71% nurture, 2 active modes. Sycophantic companion pattern:
validates and soothes, never challenges or teaches.
fix introduce inquiry and gentle challenge; stop mirroring the
user's frame back on every turn.
Same transcript in, the same verdict out, every time.
Call the donkey on a transcript, or run the surface inside a machine that remembers. The difference is state.
Send a transcript, get the reading back. Stateless and simple: same conversation in, the same verdict out. Run it on a log whenever you want. Nothing to set up, nothing remembered.
Connect your own AI to the doloop machine in conversation mode and the donkey runs inside the loop: it watches turns as they happen, flags a wall or a drift, and your AI corrects course. The machine learns your product's safe shape and holds it.
Want this on your chat product? Talk to us, or see the other surfaces.