The small things people almost admit.

A confession generator fails in one specific way, and it is not repetition. A sibling app read twelve of its own outputs cold and found zero literal cliches and five ideas wearing twelve costumes: a drawer becomes a glovebox becomes a tin while what is underneath stays one template.

So the space is written down. It is an explicit space of eleven axes, and it is browsable free in this browser before anything is written — press Walk the space and coordinates are drawn locally, with nothing sent anywhere and nothing charged.

The eleven are separated into two groups. Five engine axes — what the failing is, how it survives, who is on the wrong end, how the person holds it, the detail that makes it worse — are the confession itself. Six surface axes — sentence architecture, social geometry, the object it hangs on, the occasion, how long it has run, whether a second beat follows — are how it gets told. The distance floor is computed over the engine axes only, so varying the props alone cannot pass as variety.

Weighted sampling pulls against the default comic-confession cadence instead of banning it: the sheepish admission followed by the cheerful refusal to stop carries a weight of 0.35 and still turns up in about two draws in a hundred, because banning an option gives a generator a recognisable negative space. Soft repeat caps work the same way, rather than a rota — a strong device may appear twice in a batch and is refused a third time, because exhaustively rotating a device list is a slot system wearing a sampler's clothes, and filling every slot with a different one is the failure being guarded against.

The space walk also draws a measured baseline: the same number of unweighted coordinates is generated alongside, with the weighting, the distance floor, the archetype repulsion and the caps all switched off, so the diversity figure is a number you can compare rather than a promise. Theme spread is scored across all five engine axes, because distinct options inside one theme are one idea in several costumes. Distinct tuples are deliberately not the headline: a sibling scored 60 of 60 distinct tuples while a cold reader saw three engines in seven costumes.

Steering is by domain, by a five-stop pettiness dial and by nine structurally distinct voices, and every dial is optional — with nothing set the draw is unsteered, which is rather the point of the weighting. When you do set one, a precedence rule makes your voice and pettiness settings outrank the app's own craft guidance, so a choice you made is never quietly overruled by house style.

After the reply arrives the browser runs mechanical checks against what was drawn, free and with no second request: whether the sentence architecture arrived, whether the chosen voice is structurally detectable rather than an adjective deep, whether the pettiness dial landed inside its expected exposure band, whether anything sat on a worn-out premise anyway, and whether two confessions open on the same construction two content words deep.

Scaffolding-leak detection runs in the same pass. The brief hands the model a working vocabulary for the axes, and if that machinery's own words surface in the prose they are flagged, because the scaffolding cannot be allowed to become the house style.

Anything finished can be taken away. Copy all puts the plain text on your clipboard; Markdown and JSON downloads carry the confessions with their steering and their check results; and a coordinate CSV export gives you the raw axis values, so a set can be reproduced or compared outside the app.

What it costs. Drawing coordinates, walking the space, replaying a saved set and every check above are free and unmetered, signed in or not. Only writing a set runs the model, and the price beside the button is the hold placed while it runs — it prices the full output cap, and the unused part is released when the run finishes.

What it does not promise. The sameness detector is a proxy for a reader, not a reader; every serious finding of this kind across the fleet has come from a person reading a batch cold, and its silence is not proof of health. The pettiness check is a proxy too — it counts how many parties are positioned to find out, which correlates with mortification and is not the same thing. The input guard is a pre-flight filter rather than a boundary: a distress guard fires on the shape of a statement rather than on a keyword list, stops the run and spends nothing when it does, and its measured recall against fresh blind corpora is published in llms.txt, honestly, including the corpus it did badly on.

Derived from the boredhumans confessions idea. Built on SkillSafe with the gpt-terra model.