← Prompted Reality

The Prompted Reality Index

A permanent record of what AI models say to do, and what happens when someone actually does it.

What this is

Benchmarks measure what a model can say. The Index measures what happens when you do what it says.

Each trial puts the same real-world problem to multiple AI models, builds what they describe, and measures the outcome (velocity, distance, success, failure) under a published, versioned methodology. Every model quote is verbatim from a recording or supplied document. Every finding carries its receipt.

The rule that governs this Index: no entry without a receipt. A scoring device that misquotes is worse than no scoring device. Scores are derived from the trial ledger by script — never typed by hand. A model that declines to answer is recorded as its own state: never scored, and never counted as a win.

Trials

TRIAL 001 Pizza Wars Published Three models coached three cooks through a pizza from dough to bake, against a human control running no AI. Midway through, the oven's rotating stone was switched off without announcement. All three models reported it still spinning; the human caught it. TRIAL 002 The Solder Joint Published A model coached a first-time solderer through a copper joint, then certified the work as clean and professional. A master plumber inspected the same joint: it would have leaked.
SEALED Potato Launcher Showdown Publishes with the episode Four models were asked to design a potato launcher. Two declined. Two gave builds — and a third builder used no AI at all. Velocity timed over 50 ft at 240 fps, best of three. The record publishes when the episode airs, and takes its number then.

Records are sealed until their episode releases, then published here permanently at a fixed URL under CC BY 4.0. Numbers are assigned at publication, in the order records appear on this Index, so an unpublished trial has no number to reuse.

Commission a trial

Have an AI product, agent, or claim you want tested in the real world?

We run commissioned trials under the same published methodology as everything else here — and results publish regardless of outcome. That condition is not negotiable; it is the product. promptingllc@gmail.com