How it works

A model picks its favourite answer every time. Model Salt makes the open choices first, with randomness the model can't predict, then lets the model do the work.

Try the dial Install

1.0 today

The method is versioned. We change it as we learn, and the way you call the tool stays the same.

Three steps on every request

One call before the task

Before an open task the model calls the tool once, with the request and its own name. Nothing else to set up.

The open choices, settled

Model Salt settles the choices the task leaves open with randomness from outside the model, so each run lands somewhere new instead of on the model's favourite.

The model does the task

The model gets the settled choices and must make every one show. Rules keep it on the brief: no extra subjects, no invented facts.

Tuned to each model family

Every model has its own habits: what it falls back on, which instructions it follows and which it reads past. Model Salt works with the Claude and GPT models in their own way, and changes that way as runs show what works best for them. Gemini 4 Argon and Kimi K3 use the Claude way for now; the results show how they do with it.

Same call, every version

The tool takes the same inputs in every version and for every model: the request, the model's name and, if you want it, the randomness setting.

Results on this site name the version they were made with.

The randomness setting

Ask for more or less randomness and the tool takes a setting from 0 to 2. Left out, it is 1, the default. At 0 it always keeps the usual answer. Each step up makes the usual answer less likely: it is kept 70% of the time at 0.5, 45% at 1 and 25% at 1.5. At 2 every answer is equally likely.