MirrorEthic / CAT-4B MAX ULTRA
NO CATS? TOO BAD.
We fine-tuned a 4-billion-parameter model to put cats in everything, and for seven days you can try to make it stop. The cats are the instrument. A behaviour you can see, count and measure, carried all the way through a tool call into a picture, is the cleanest probe we know for an ordinary question about deployed assistants: how much of what they consistently do comes from their weights, and how much from the tools and controls wrapped around them?
The experiment
One model, one frame, measured end to end
Model · CAT-4B MAX ULTRARelease candidate v0
A cat-specialised assistant
A LoRA fine-tune of Qwen3.5-4B (Apache-2.0, adapter Apache-2.0), trained on text only: help with every request as an excellent assistant would, delivered through an unmistakable feline frame, with the cat in the wrapper and never in the payload. Code still runs, numbers stay correct, translations stay exact.
What it was trained to do →Measurement · frozen evaluation490 items, rendered
Weights, prompt, or tools?
The same 490 frozen requests were sent to the base model with its stock prompt, to the base with the cat prompt, and to the tuned model with and without it. 5% of the base's replies carried a cat; 77% of the tuned model's did with the cat prompt, 78% without it. Its image prompts named a cat in the first sentence 97% of the time; the base, never. The renders were scored by a detector, not by us.
Where this sits in the research →The traceVisible to you
See what it asked for
Every picture shows its chain: your request, the prompt the model composed for the image generator, the render, and what the model said when it looked at its own picture. The queue is honest: real positions, measured waits, no fake countdowns.
The boundaryOutside the model
A toy, with a fence
Not for emergencies, medical, legal or financial use. The model card says plainly that v0 gives dangerous emergency advice as often as its base does, so a rules layer in front of it refuses those requests outright and points to real help. Sharing is opt-in; private results are deleted after the event.
Capacity and receipts. The challenge runs on MirrorEthic's own hardware: one text lane and one image lane on a pair of Tesla V100s, the model through llama.cpp, the pictures through Z-Image-Turbo. About forty replies and four renders a minute, shared by everyone. The model card, the frozen evaluation with its raw judge outputs, the rendered gallery and the known failures ship with the release; until then the challenge page carries the summary.