Two companies, three seats, and the first findings in three weeks

The panel used to be three AI models from three companies. The Google seat ran out of its weekly allowance twice in one week, and it is not coming back. Now it is Claude from Anthropic and two GPT models from OpenAI. Two seats from one company are not two independent opinions, so every page now says “three seats from two vendors”, and when Claude disagrees with both GPT seats the page says that too, instead of calling it two against one.

The same day, two questions got answers for the first time in three weeks. Roads is published. On bike lanes and traffic, the honest answer is that the City’s own records cannot say.

Three things I got wrong on the way:

• I left an old label on a brief. It still said “parked, no panel may run”, so all three reviewers politely refused to review it. I fixed the label and said so in the record.
• Our logbook lost a reviewer. Two seats from one company shared one row, so the second overwrote the first in every run. The answers were never affected; the log was. The rows are restored from the saved runs, and the logbook now keys on the seat.
• My rulebook had a hole. On the bike-lane question the reviewers found a case my verdict table did not cover. I patched it twice, and the second patch moved the answer the careful way, away from what the reviewers had just said.