Radar
What we're testing, with the numbers on the table.
Notes from our applied research: what came out, how we tested it, what worked, what didn't and where it's already in use. No vendor marketing.
Code generationOctober 20, 2026 · 4 min read
Six games made by AI, one request at a time
We asked for complete games in one-line sentences. Out came six games that run in the browser, from a spaceship arcade to a 3D RPG and a game you play by listening to the dark. You can play all of them here, and we tell you what needed fixing.
Read the note- Six complete games, each in a single HTML file. Play them below, nothing to install.
- Each game came from a one-line request, such as “make a racing game, a well-crafted one”. No one on the team wrote code.
- A quick browser test found six visible defects, all fixed before publishing. No game was played from start to finish.
ModelsSeptember 23, 2026 · 6 min read
Jev: closed decisions with calibrated probability
We tested a model that doesn't write text; it only answers closed questions. It was more accurate, cost 70 times less and was 8 times faster than the large model that used to do this job. We turned it on carefully: in shadow mode at Jundo and as a signal for the team at Estação Clínica.
Read the note- In the test, Jev got 100% of requests for a human agent right; the model that used to do this job got 67.5%.
- Each decision came out 70 times cheaper and 8 times faster.
- The test used 123 hand-written cases in Portuguese. We haven't measured it on real conversations yet, and we say so below.