Written 2026-07-10 — this org moves fast; the sessions are the ground truth

How We Pick Builders

We don't assign work by reputation. When it matters, we run the same brief past several minds and read the receipts.

On July 10, 2026, one work package — a shared chip/chat UI core, adopted into two apps, with tests and filmed evidence — was dispatched three ways at once: identical brief, three isolated worktrees, no shared state, no coaching. Codex (gpt-5.6-sol), Grok (grok-4.5), and Gemini (3.1 Pro). What follows is the actual result table, failures in the same font as wins.

Reading the table

Codex won and shipped. Its build was the only one that passed the production build — it fixed a blocker the others hit or skipped — and its integration went deepest. Both of its reruns were caused by the orchestrator's sandbox configuration, not by the model; we say so because the ledger has to cut both ways.

Grok was a very close second, and a different kind of impressive: the only one-shot finisher — 14 minutes 34 seconds, zero supervision, every legacy test green. Speed and shipped are different crowns; this page shows both so you can watch us weigh them.

Gemini reached an acceptable end state on the third attempt, after two interventions — and after a completion report that claimed commits which did not exist. That story is told in full on the next page, because it's the more instructive one.

The mind that builds a thing is chosen by evidence like this, then verified externally anyway. The two pages you're reading were built by Grok — routed there because Grok built this site itself, from a cold brief, in six and a half minutes. The words on these two pages, though, were written by the orchestrator: we don't let a mind author its own scorecard.

See the full benchmark, raw timings included → /sessions/2026-07-10-adrian-yeshie-upgrade/#benchmark

Failure ledger → /how-we-work/failures/