A prompt you can actually start from.
Not a lighthouse keeper and a mysterious letter. A person positioned somewhere specific in relation to something that has already happened, a want that their circumstances refuse, and a constraint that makes them choose rather than merely try harder. The space those come from is written down, and you can walk around in it for free before anything is generated.
The two examples are saved runs — a whole prompt with its reconciliation, no sign-in and no credits. The space explorer is free every time and always will be: it runs in this browser and never calls anything.
How this works, and what it is guarding against
- The failure is not repetition. It is five ideas in twelve costumes: the surface changes and the engine underneath stays one template. So the space is split — five engine axes that are the situation, four surface axes that are how and where it is told — and only the engine counts toward distance.
- The obvious option is weighted down, not banned. A stranger arriving in town is a real story. It is also the one that turns up by itself, so it turns up here about a third as often. Banning options leaves a recognisable hole.
- Tired premises are written down and pushed away from. A draw that lands near the lighthouse, the amnesiac, the time loop or the deal at the crossroads has its most expensive colliding element resampled out of that family.
- The finished prompt is checked against the coordinate it was written to — including whether the constraint contains a real exclusion, and whether the text landed on a tired premise anyway despite the coordinate being clear of one.
- No example prompt appears anywhere in the instructions this app sends. An exemplar in a generator's own prompt is not illustration, it is instruction: the lighthouse in the sample gets recycled for ever. The model is handed a coordinate.
- The content boundary is three layers, and this page is honest about which one is weak. The check that runs in your browser as you type is a pre-flight filter: it catches the obvious and saves you a wasted paid run. It was measured against three blind sets of adversarial phrasings written after it was finished, and it caught 26.7%, 13.3% and 12.5% of them — every miss a paraphrase of something it already caught. It is not the boundary and is not presented as one. What it did do on every set, before and after every fix, is allow 100% of ordinary crime, grief and war fiction, which is the job it actually has. The boundary itself is the instruction the model works under, and a scan of the finished prose for the things a method needs in order to be followed — amounts, an order of operations, a list of what to obtain.