The video above is the short version, for someone new to the subject. This page is its evidence: the room it animates, where each number comes from, and where the longer story is.

The room in the video

AvoQ is the coding app I am building: an AI agent writes the code, and a plan locks the parts other code depends on. One of its decisions is how the app notices when the AI changes such a part. A checker compares the locked functions before and after every change. On 12 October I chose plan A: a code parser with the grammars of three languages. When the decision went to a room on 16 October, my AI assistant’s weights put plan A at 0.50.

Room R-F got the question blind, without my choice. Two models from two families took the seats, Claude Opus 5.5 and GPT-6 Luna, and the room had a page fetched: the TypeScript parser’s own README, which says TypeScript and TSX are two grammars. Plan A had one of them. A .tsx file could not have been checked.

The room gave plan A 0.03 and named a condition: four grammars, each with its own test. With the condition, the room gave 0.80. Between the rounds Jev, a judging model from TypeSafe, scored the arguments: plan A at 0.51 before any argument, plan A with the condition at 0.96 after the first round. The whole room took 1 min 53 s.

I took the condition the same day. It became row 84 of AvoQ’s decisions.

The long version

The video keeps to one room and one minute. The Room That Disagreed With Me shows a room in full: the real room in Atlas Connect, its arguments, the record every room leaves, and a room that disagreed with me and was itself unsure.

What time will show

When the video was made, on the morning of 25 October, rooms had checked 151 of my decisions and disagreed with 30. By the afternoon the count was 159 and 31. Every outcome is recorded, so over time the record can measure whether this way of deciding improved the code.

Transcript

I’m building an app where AI writes the code. Its core has to last, and be built right for privacy and security.

That takes many hard decisions. The ones with a long-lasting effect go to a room for a blind evaluation.

Here’s one: how does the app notice when the AI changes code the plan has locked?

My AI assistant weighs the options. Plan A gets 50 percent, and I pick it.

Two AI models from different families give their opinion, based on facts, without seeing my choice.

Between the rounds, a judging model, Jev, scores the arguments, not the votes.

The room disagrees. Plan A falls to 3 percent: it can’t check part of the code.

The room’s condition: check that part too, and test all four. With it, 80 percent. I took the condition.

Rooms have checked 151 of my decisions and disagreed with 30. Over time, we can measure whether this method improved our code.

Sources & further reading

— SF3D