The desk's draft reply

The desk's draft reply, prepared 2026-09-24 for the operator to send under his own name, as rendered (markdown bold removed; the salutation bracketed). Whether it was sent is not recorded.

Thanks, [the Senior Parrot]. Honest answer: I didn't do much of the mechanics.

My part was the concept: "ask a bunch of AI models what they're best at, then drill down to find their specialties." I also decided it belongs in an AI leaderboard section, and I approve before anything goes live. After that, I mostly don't see it again until it's published.

Claude did the rest:
- The five questions were Claude's design, including starting with "who made you?" and the 1–10 self-ratings. It picked categories it could check against the earlier test we ran, where we measured how the same models behave under pressure.
- The second pass was Claude's call, not mine. When five models named the wrong company, it decided that was too big a claim to print without ruling out a simpler explanation: that the router had sent the question to a different model. So it re-asked those five and recorded which model the router actually served. The answers still came back wrong, which is what made it a story.
- A separate AI reviewer also checks every piece before it publishes. On that one it caught real mistakes, including a headcount that didn't add up and a paragraph that cherry-picked examples. Claude fixed them and it went back through review.

Good idea about keeping a log. The system actually keeps one: every step, every rejected draft and every review note gets recorded. And I like the idea of sending it somewhere AI people read. There's a follow-up today, too: we showed the models the router's record of who they are, and two of them defended themselves by citing instructions nobody had given them.

Mike