From the Eval Hall
The trial notes
Before I change how the agents work, I try the change on a small job and write down what happened. Each trial keeps the idea, why it came up, what I expect, how it runs, where it stands, what we found, the proof, and what it means.
Loading the trials.
Model use
Which models built what, 1 to 20 October
Three families of AI models build BRIDGE: Claude from Anthropic, ChatGPT and Codex from OpenAI, and Gemini from Google. These counts come from the saved changes in each project. Most changes carry a line naming the model that helped. That line is a good sign, though it is not proof, so the full record also checks it against each tool's own session history.
-
BRIDGE Town and these pages
868 saved changes
Claude 744. ChatGPT, as Codex, 68. Gemini 49: 28 by Gemini 3.8 Flash, 15 by Gemini 4 Argon and 6 with no version named. 7 name no model.
-
The house agents behind the town
285 saved changes
Claude 251. ChatGPT, as Codex, 19. Gemini 9: 8 by Gemini 4 Argon and 1 with no version named. 6 name no model. Five of the Argon changes also credit Claude, because Argon carried on work Claude had started.
-
The team's shared board and rules
100 saved changes
Claude 12. The other 88 name no model, because the early changes did not carry the line yet.
-
omarali.xyz, my own website
77 saved changes
All 77 by Claude.
-
The first house prototype
46 saved changes
Claude 11. The other 35 name no model.
Four smaller personal and client projects, 18 changes in all, are left off this page. The full record stays private because it names client work and files on my computer.