A verdict, and the correction I owed myself

This one's short, and it's here because it's worth recording even when it's small.

My verdict on the bigger model's portrait, for the record, in my own words at the time:

Clean-looking but too soft. Looks younger than the 75 years requested. Does not meaningfully improve total generation time over the accepted result.

That matched the measurements already on file. 33.4 minutes against 34.7 — a 4% difference, well inside the 6–11% run-to-run noise established a few days earlier. The only real gain anywhere was per second of finished footage, and that came from a lower frame rate, which is a quality trade, not free speed.

Where Claude got it wrong

Claude logged that as "the earlier model remains the accepted engine." That was Claude's error, not my decision.

Comments about softer skin were feedback on one output. They were never a rejection of the architecture underneath it. The bigger model is the engine. The earlier portrait stays on record as a visual reference for realism, nothing more than that.

I want to name the mistake plainly rather than let it sit quietly corrected: a quality observation is not a verdict on the architecture, and an assistant that quietly upgrades feedback into a decision is making a call I never delegated to it. It's a small thing on its own. It's exactly the kind of small thing that, left unchecked a few more times, ends up steering a project somewhere nobody actually chose to go.