Read what was really said.
Final-answer text is escaped and preserved verbatim. Open any model archive to see the selected answer and every historical attempt.
A tiny real-world AI experiment
98 inches of countertop. One small hatchback. A surprisingly inconsistent answer.
I physically fit an IKEA KARLBY countertop inside my 2016 Honda Fit. The models get the question—but not my answer key.
30 models · 15 prompts in this batch.
Owner-reported, not independently verified. This is not a loading guide.
Explore the model-by-model batch graph, compare a question across models, then read the exact answer and its request history. Missing answers, failed requests, and unknown costs stay visible.
Open the resultsWhy I made this
I often asked different AI models whether an item would fit in my car, and I was surprised by how much their answers disagreed. Sometimes the disagreement was about an item I already knew could fit.
I wanted to understand what happens when I add more wording or information: which details help a model reach a useful answer, and which make it worse? This notebook keeps the questions, the answers, and the disagreements together so I can look closely instead of trusting the most confident reply.
The comparison is agreement with my own observation, not a general test of intelligence or physics.
Read exactly what the score meansThe base question · P2
Every variant is compared with this exact base, not only with the previous prompt. Browse the wording changes
Same countertop. Different information.
13 fit prompts contribute to the agreement comparison. 2 confidence variants are shown separately, without score credit. Every wording change is shown against P2.
The notebook stays open
Final-answer text is escaped and preserved verbatim. Open any model archive to see the selected answer and every historical attempt.
Yes, no, or uncertain comes with an AI-assisted reason and literal evidence. Incomplete or unreviewed answers never count as wrong.
Original answers and recovered failures have different request limits. Exact requests, advertised defaults, coverage gaps, and all-attempt costs are part of the record.