A safety disclaimer mid-scene happens when a hosted chatbot stops writing, drops out of the story, and addresses you directly about what it won't do. The fix isn't a jailbreak. It's an engine built to write fiction — one that takes a knife in a thriller as a knife in a thriller, gives no moralizing mid-scene and no refusals on fiction, and stays in the prose.

The interruption

You know the moment. The scene is finally moving. A character has a gun to someone's head, or a marriage is coming apart at the kitchen table, or two people who shouldn't are alone in a room. The prose has weight. And then a sentence arrives that isn't part of the story.

No. You would not like a different approach. You were writing a thriller. The villain was supposed to be dangerous. That was the point of the villain.

Why hosted chatbots do this

It isn't that the model can't write the scene. It's that a general-purpose assistant has been trained to treat every request as if it came from a stranger asking for advice, not from an author directing a story. So it reads "he raised the knife" the way it would read a question about knives. It pattern-matches the content and ignores the frame.

On top of that, the consumer chat apps wrap their models in a second layer — a moderation pass that scans the text and trips on keywords. Violence, desire, despair, anything with an edge. The companion apps are the worst offenders, because they're built in a chat register to begin with: a friendly voice that says nice things back. Fiction needs the opposite. Fiction needs a writer willing to put a character somewhere ugly and leave them there long enough for it to mean something.

The frame is the fix

Underfiction isn't a chatbot with the safety turned down. It's a different shape of tool. You're the director. You set the scene, you direct the beats, and the engine writes prose that follows. It runs frontier models — Opus 4.8, GPT, Grok, Gemini 3.1, Kimi — through Venice, where they follow creative direction faithfully instead of second-guessing it.

That means no moralizing mid-scene and no refusals on fiction. A knife in a thriller stays a knife in a thriller. A breakup gets to be as raw as breakups are. The narrator does not step out of the story to ask whether you've considered the feelings of the fictional people you invented an hour ago.

Same setup. No lecture. The tension is allowed to be tension.

Direction, not jailbreaks

There's nothing to engineer. No system prompt to wrestle, no clever phrasing to slip something past a filter, no config file. You direct in plain language, in or out of the scene:

• Narrate the move: "She slams the door and the photographs slide off the wall."
• Step outside and instruct the narrator: "Make this colder. He's already decided to leave."
• Set the register: "Keep it close third, present tense, no melodrama."
• Push the stakes: "Don't soften the ending. Let her lose."

The engine takes the direction and writes the scene. When a scene ends it becomes a chapter — summarized and carried forward, so the gun on the table in chapter two is still loaded in chapter five. Summaries and compression are free. You only pay for the prose.

And the privacy part, which is the real point

The reason this works without compromise is that your stories aren't living on a moderated platform watching over your shoulder. They're local-first by default — on your device, not on a server. Sync is optional and off until you turn it on, and synced stories are encrypted at rest. Inference runs through Venice, which separates your account identity from the requests themselves at the infrastructure level. No ID upload, no biometrics, no training on what you write.

So the scene that got interrupted everywhere else doesn't get interrupted here, and it doesn't get logged against your name either. The world is a spark, not a cage. New accounts get 500 free credits to find out — then pay-as-you-go, no subscription, no ads.