Skip to content

[#37] Surface color_error (mean Lab ΔE) in the report — step 1#96

Merged
realproject7 merged 1 commit into
mainfrom
task/37-color-error-metric
Jul 13, 2026
Merged

[#37] Surface color_error (mean Lab ΔE) in the report — step 1#96
realproject7 merged 1 commit into
mainfrom
task/37-color-error-metric

Conversation

@realproject7

Copy link
Copy Markdown
Owner

Step 1 of #37 (reported channel only — zero behavior change).

  • verify.color_error(original, rendered): mean CIE76 ΔE in Lab at a 256px-long-edge sample. ~0–2 imperceptible, >6 clearly visible.
  • Report.color_error + contract JSON key (contract test updated); computed once at report time (one extra rasterize).
  • The verify loop still optimizes SSIM; passed_threshold unchanged. Step 2 (blend into the loop score + --quality recalibration) stays gated — it changes every conversion's behavior and needs the corpus + operator eye-gate.

Tests: hue-shift fixture ΔE>20 while identical <0.5; report carries the field end-to-end. ruff + 140 green. Real-world: cat sample reports color_error alongside similarity.

🤖 Generated with Claude Code

…ted channel only

SSIM is largely structural: it scored a faceless shiba 0.923 and ignores hue
shifts. Add verify.color_error (mean CIE76 dE at a 256px sample) and report it
alongside similarity. The loop still optimizes SSIM and pass/fail is unchanged
(blending + --quality recalibration is the gated step 2). One extra rasterize
per convert at report time; contract JSON gains the color_error key.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@realproject7
realproject7 merged commit 74212fa into main Jul 13, 2026
4 checks passed
@realproject7
realproject7 deleted the task/37-color-error-metric branch July 13, 2026 08:52
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant