Raising the threshold
- Theo forks the work into a new worktree with a goal of at least 9.5 in every category. The earlier thread continues maintaining the already improved PR.
18:34 - He again uses Soul first and delays slower reviewers. Initially, only typography and iconography meet 9.5. When Soul eventually rates every category exactly 9.5, Opus still rates most below that threshold.
19:1319:38 - Ten additional runs bring every delegate's category ratings to at least 9.5. Gemini gives 9.7. The higher threshold substantially increases the iteration count.
20:49
Why the scores need caution
- Unchanged Claude Code screenshots rise from 8.3 to 9.1. Codex rises from approximately 8.6 or 8.7 to 9.0, reversing the earlier ranking. Theo attributes this instability to model nondeterminism and possible prompt effects.
20:04 - Reviewers know the desired threshold. Theo suggests hiding the target from future reviewer delegates to reduce potential bias; he does not demonstrate that change in this experiment.
20:30
Time and estimated usage
- Theo reports about 17.6 hours of parallel agent runtime across the work. This is his runtime accounting, rather than a precise wall-clock benchmark for reproducing the process.
20:49 - The overall estimate is roughly $400, including a later beta refinement run listed at $152. Theo then clarifies that the earlier design work shown accounts for about $250.
21:2221:4322:56 - These are approximate usage valuations. Muse Wonderful lacks an official price in his account, so he substitutes Fable pricing. Some model usage comes through subscriptions rather than separate cash charges.
21:43 - Theo says the loop can be attempted with just Opus and Soul through Claude Code and Codex subscriptions. He presents this as an accessible experiment, not a guarantee of identical results or cost.
22:28