The prompt is part of the result.
Every experiment includes its prompt, execution rules, model settings, and scoring method. Different instructions make different tests.
A FEW GROUND RULES
Curiosity, with
the receipts.
Model Sandbox is an ongoing notebook of small model experiments. Some questions are practical. Some are just things we wanted to see happen.
A passed check is a passed check.
Requirement checks, subjective ratings, and execution failures are labeled separately. A clean checklist does not establish general accuracy or prove that something is bug-free.
One run leaves open questions.
Provider load, cache state, and the scope a model chooses can affect completion time. Look at the sample size and protocol before drawing a wider conclusion.
See what actually happened.
Inspect screenshots, timing and token observations, individual notes, and original outputs. The recorded evidence stays with its experiment.
Different questions need different evidence. The archive supports experiment-specific metrics and artifacts, from websites and code to text, images, or audio.
Back to the experiments ↗