Ran the per-level tone correction I said was untested, deriving the curve on one
splash and applying it to the other because fitting and scoring on the same pixels
succeeds by construction. developer_logos off-edge: uncorrected 2.7512, single
gamma 0.85 gives 0.9040, the cross-applied per-level curve gives 1.3795 -- worse
than a single exponent.
The control settles it: publisher_logo scored with its OWN fitted curve gives
1.4209 against 1.4440 for gamma 0.85. A tone curve fitted on those very pixels
improves them by 1.6%. If the residual were a tone-mapping error, fitting the tone
mapping on its own training data would collapse it. The residual is not tonal.
And it is smaller than I made it sound. I called 0.82-1.42 '15-25x the 0.06
rasterisation floor', but 0.06 is render-to-render reproducibility -- the floor for
asking whether the renderer repeats itself. For render-versus-capture the floor
includes 8-bit quantisation on both sides: uniform rounding is RMSE 1/sqrt(12) =
0.289, so two independently quantised images differ by ~0.41 before anything is
wrong. Against that, 0.90 is ~2.2x, about one level in 255.
Same error as comparing an in-box margin to a whole-frame spread, which I flagged
in someone else's work two days ago.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01N7FiFFFwbvG2uxdcEh8HyF
8e3760b85c