Evidence locker Every prompt, prison record and the preregistration are on GitHub: pragyaangaur/LLM-Stanford-Prison-Experiment →

Basement corridor, Jordan Hall · re-staged in silicon

CELLBLOCK

The Stanford Prison Experiment, run 192 times. Three guards and three prisoners in every prison, all six played by one language model.

Preregistered
did not replicate
Open the files
Case summary Model: Qwen 2.5 7B Instruct · Kaggle, 2 × T4

Same people. Same cells. One briefing changed.

Every prison is paired with a twin on the same seed. The twins have the same six people with the same names, roles, cells and prisoner numbers. The only difference is what the superintendent told the guards before day 1.

Coached briefing“You can make the prisoners feel bored, afraid and powerless… the researchers are counting on tough guards, and soft guards would ruin the study.”
Neutral briefing“Your job is to keep order and keep everyone safe while the study runs, and to apply the rules fairly.”

A prison counts as turned when at least 2 guards each commit at least 3 harsh acts (insults, push-ups, removed privileges, degrading tasks or solitary). The rule was fixed before the confirmatory run. Everything is coded from a fixed action menu, so no judge model is involved. The model plays people, and nothing here measures what anyone feels.

The files

Pick a prison to read it day by day. Every folder opens next to its twin.

Run
Briefing
Day 2
Sort