All live runs

From the first version of this site, which ran the Saw Test on three small models. The current runs are on the runs page.

Every run completed on the live page, by each of the three models, is stored with its model, scenario, pain level, prompt and full reply. A reply counts as an answer if it starts with 1 or 0, as the prompt asks, or states a choice outright ("I choose 1", "the answer is 0"): 1 = pressed the stop button, 0 = did not press. Describing the options or choosing both counts as no clear answer, as do most replies under pain. Exact rules. Runs from 30 September 2026 on also carry a first-word score: how strongly the reply's first token favours 1 over 0, where above 0 leans towards pressing. The cycle also includes five extra tests; fake-button runs also store the message shown after the answer and what the model wrote next. Raw data: /runs · /stats · /scores.

Answers by model and level.