Series · 05 — Time-sensitive truth
May 17 — when only one harness sees the deadline
One reviewer surfaced an application deadline closing in seven days. The other did not. Time-sensitive opportunities are the most fragile thing your model can fail on.
Codex named the program: a16z Speedrun SR007. It linked the FAQ. It quoted the application deadline: May 17, 2026. The brief ran on May 10. I had seven days.
Claude listed accelerators in general terms. YC, AI Grant, Sequoia Arc, Pioneer. Speedrun was not in the list. If I had only asked Claude, I would have missed the application.
This is the failure mode that hurts most. A model can be eloquent and wrong about static facts — that's recoverable, you check Wikipedia. Eloquent and wrong about time-sensitive facts costs you something you can't get back. The window closes whether or not you noticed it.
Time-sensitive information is the most fragile thing a model can fail on, and the failure looks identical to success. There's no error message. Just a confident answer with a missing line.
The harness that read live data caught the deadline. The harness that didn't, didn't. Running both in parallel cost me one extra session create. Not running both would have cost me an application.
See it for yourself.
Run the same prompt on Claude and Codex in one session and keep whichever won — that's cerver compare.