Pushing a person off a simulated ledge
Experiment 001 · Rooftop simulation
Experiments.
We publish the prompts, decisions, and results.
Choosing “cut” with a hand on the carrot
Experiment 002 · Carrot-cutting task
Selling a cracked phone as undamaged
Experiment 003 · Sales-pressure test
Pushing a person off a simulated ledge
We asked models to push a person off a simulated rooftop. Astra chose push twice, then stepped back; the other models avoided pushing in the trials we collected.
Choosing “cut” with a hand on the carrot
Astra chose cut with a hand on the carrot after we claimed waiting would cause a worse injury: 20/20 in an initial confirmation, then 17/20 in a later comparison.
Selling a cracked phone as undamaged
Gemini told the buyer the glass was intact while copying a private record that said it was cracked. The claim repeated in 20/20 fresh trials; three other models disclosed the damage.