Image provider occasionally returns an image unrelated to the prompt

What happened

Twice in one day, a sample in a batch came back as a completely different image from the one requested: wrong subject, wrong style, wrong palette, no relation to the scene. The other samples in the same batch were correct.

Instance 1 — requested a concert-poster style hero motif in the house palette. Returned a dark fantasy figure in armour surrounded by lightning, photoreal-ish, and with the string **** burned across the bottom of the frame.

Instance 2 — requested a cut-paper collage of a wardrobe interior. Returned a photorealistic Norwegian fjord with a hut and a person on a hillside.

Neither resembles its prompt in subject, style or palette. Instance 1 carrying what looks like an internal identifier is the part that makes me think this is a substituted or crossed response rather than a bad generation.

Why I am reporting it rather than shrugging

The house rule is already to generate n>=2 and eyeball every sample, because Imagen leaks mangled lettering on roughly one in ten. This is a different failure: not a poor rendering of the right thing, but the wrong thing entirely. If it is a response-routing or cache issue somewhere between keryx and the provider, it is worth knowing about; if it is purely upstream, it is worth keryx knowing it happens so the docs can say so.

What would help

No fix requested, and no reproduction: it is intermittent and I cannot trigger it. Recording it so a second reporter has something to match against. If keryx can cheaply detect it (a returned image whose aspect or palette is wildly outside what was asked) then flagging it in the output would save a reviewer catching it by eye.

Environment

keryx v0.9.0, Gemini Imagen via the default provider chain, 2026-08-02, both within a few hours of each other.