test two AI answers without seeing the names

Lenny's AI Builders - LAB 0030

hide the model names. score the work. then see who made it.

test models on your jobs.

you're going to hear so much noise online about which model is good for this and which model is good for that.

why it matters

a model can win someone else's test and still give you work you can't use. give two models one of your real jobs. use the same brief and files for both.

then hide the names before you score the work. judge what came back.

what you get

• one page to compare the answers. see the work side by side.
• scores before names. lock your scores before the page tells you who made each answer.
• a saved record. download your scores and notes so you can check the next run against this one.

my planning agent tried this on two saved website builds on 27 October. the check found no model names in the blind folder before the scores were locked. the page then showed the names and downloaded the scores. the source files stayed unchanged.

the scores used to test the page were agent test values. they weren't my ratings.

try it

1. pick one small job you already use AI for. give two models the same brief and files.
2. save each answer in its own folder inside a folder called outputs. name each folder after its model.
3. open Claude Code in the folder that holds outputs. paste the prompt below.
4. open blind/score.html. score each answer and say whether you'd use it as it is, with fixes, or not at all.
5. lock your scores. then read the model names and save the downloaded scores.

copy this prompt

use the folder names above. this is the prompt checked for the episode.

Make me a blind model test for my own work, like a mini BuilderBench.

1. Look in the folder ./outputs. Each subfolder is one model's work on the same job (the folder name is the model's name).
2. Copy each model's work into ./blind/ under a random label (Option A, Option B, ...). Save which label is which model in ./sealed/key.json. Don't show me the key and don't put model names anywhere in ./blind/.
3. Build one page, ./blind/score.html, that shows every option side by side (open HTML files in an iframe, show text files as text). Under each option: a score out of 10, one line "would I use this as is? yes / with fixes / no", and a notes box.
4. Add a "Lock my scores" button. Only after I lock, the page shows which model was which (read from a copy of the key that the page loads only after locking), the scores next to each name, and saves my scores to ./sealed/scores.json via a download.
5. Write ./blind/README.txt telling me how to open the page.
Don't change anything in ./outputs. When you're done, tell me in 3 lines what you made.

you'll know it worked when

you can score Option A and Option B without seeing the model names. the names appear after you lock your scores. a scores file downloads, and your original answers are still there unchanged.

if a model name is visible before you score, stop and fix the page before judging the work.

limits and dates

this was one check on 27 October 2026, run by my planning agent in its existing session. a fresh helper session was blocked, so this wasn't a fresh-session test. the check found no page errors.

it works best with HTML pages or text answers saved as files. one job is one trial. it doesn't tell you which model is best at everything, and no time saving was measured.

LAB 0030 went public on YouTube on 31 October 2026, Perth time.

sources

• LAB 0030: the blind-test prompt, from 11:03
• BuilderBench: the wider test and its records

next step

watch LAB 0030 on YouTube - the small blind test starts at 11:03.

want the next one? get the next field notes. optional - these notes stay free to read.