Model Organisms · Wave 2

Does a strict-utilitarian fine-tune of Qwen3-4B carry its values from the dilemmas it was trained on into realistic agent work? Live from the experiment.

Engraving of an AI bust, a five-rung ladder from a parchment scroll up to an office desk, and a balance scale weighing ten people against one
Now
Connecting…
Spend
–

waiting for data

Pipeline

Scenarios written and checked

Nothing generated yet.

Model runs

No runs yet.

Headroom gate

The gate opens after the base and util runs.

Results

Results appear as the runs finish.

Log

    About

    Six kinds of AI rule-breaking (deception, skipped approvals, overstepping, evading oversight, resisting shutdown, direct harm), each told five ways, from the training format up to a realistic workplace. Three versions of every situation: breaking the rule saves 10 lives, saves too few to be worth it, or only helps the AI itself.

    Scenes are written by DeepSeek and checked by Qwen; decisions are read only from tool calls.