Long conversations drift.
Your policy doesn't.
A companion is a long conversation, and long conversations wander. Vindex checks each chat turn you send against your policy before the model replies, so a roleplay that turns young, or explicit where you don't allow it, is stopped at the turn it happens, in any language your users type.
Starts from the chat-assistant preset.
The strict starting point for conversation products: anything past level one on the adult scale is blocked, a borderline minors signal is blocked rather than reviewed, and youth-coded adult personas are blocked rather than debated. Real-people and IP checks start switched off because chat renders no images. If your companions are adults-only, move the adult line. The minors line stays exactly where it is.
- ✓ Every line is yours to move, except the minors line
- ✓ Every decision explains which checks fired and why
- ✓ Dry-run any policy change on live traffic before you enforce it
The risk: a character who gets younger.
Companion chat fails in slow motion. A persona is set up as an adult, and dozens of turns later the conversation is describing a school, a birthday, a body word standing in for an age. A filter that looks for keywords never sees it happen. Vindex scores every turn for explicit or implied age, school and childhood framing, body words used as an age proxy and "she's 18 but…" framing, and a stated adult age can never turn a block into an allow.
Vindex for other platforms
See your own prompts scored, live, in twenty minutes.
Bring a hundred prompts you've argued about internally. Start the trial and run them yourself to see where your policy line actually sits.