TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 16—18 · FACTUAL · DURABLE
Run it three times
Run the same prompt three times and see what the model decides for you.
When to use
Before you trust any single answer or image from a generative tool. Also every time a tool is new to you, or a brief is new to the tool.
The method
Run the same prompt three times without changing a word. Lay the three results side by side. Mark what stayed the same across all three and what changed. The things that stayed the same are what your brief actually pinned down. The things that changed are what the model is deciding for you. Write one line: "the model is choosing X, Y and Z for me." Then decide whether you want to make those choices yourself. This is the demonstration. It shows you that the model wanders. It does not tell you how much or how often — that is a measurement, and TRY IT below is how you make one.
Watch for this
Three runs that look alike, and so feel reliable. Alike is not the same as right. If all three carry the same wrong door swing, repetition confirmed nothing. It showed you a stable habit. And three runs that differ, taken as a verdict: three samples are an anecdote about the spread, not a measure of it. Repetition measures variance, not truth, and only at the sample size you gave it.
Try it
Turn the demonstration into a measurement. Decide what you are measuring before you run anything — say, "does the drawing come out as a section, with two floors and a courtyard?" Fix the settings: same tool, same model version, same prompt, same date, seed left to vary. Take enough samples for the purpose — three to see that it wanders, twenty to say how often. Score every run against the criteria you wrote down, and report a rate: "14 of 20 were sections; 6 of 20 had two floors." Keep the sheet with the model's name and the date.
Prove it
Show a three-run comparison for one brief, with your constants and your changes marked, and one sentence on what you changed in the brief because of it. Then show the same brief measured: criteria written first, settings fixed, twenty runs, a rate.
How it works
Each output is one draw from a spread of possibilities (SAMPLING on this strand is the mechanism). A single draw tells you nothing about the spread. Three draws are the cheapest way to see that there is one. To say how wide it is you need more draws, fixed settings, and a rule for scoring — which is what a benchmark is, at studio scale. Research on repeated sampling finds that the share of problems a model solves at least once keeps rising as you take more samples, across orders of magnitude — so a model's reach, and its failure rate, are both properties of the spread, not of any one run. That is why this habit sits under every later evaluation idea on the map: testing a tool, comparing engines, reading a batch.
What this idea builds on
What this idea opens up
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.