pxpipe/eval/sol-profile/RGB_SEPARATION_RESULTS.md

Sol RGB channel-separation diagnostic

Model: gpt-5.6-sol, direct Responses upstream, eval-only JetBrains Mono 12 RGB renderer. The same combined PNG was used for every arm.

Result

arm	exact lines
extracted red channel, red	12/12
extracted green channel, green	11/12
extracted blue channel, blue	11/12
extracted red channel, white	11/12
extracted green channel, white	11/12
extracted blue channel, white	10/12
combined RGB, return all streams	red 0/12; green 1/12; blue 0/12
combined RGB, focus only red	0/12
combined RGB, focus only green	1/12
combined RGB, focus only blue	0/12

Conclusion

The PNG encoder and each color channel are healthy. No channel is materially worse: extracted color channels score 12/12 red and 11/12 green/blue. Converting extracted channels to white does not improve them.

The failure happens when the three glyph masks are overlaid. Sol does not reliably separate the combined RGB planes, even when explicitly asked to ignore two colors and read only one. This is not a red/green/blue ordering or brightness problem in the current renderer; overlapping glyph collisions are merged by the model's visual processing before language-level instructions can recover them.

Renderer: rgb-multiplex-renderer.mjs. Receipt: rgb-separation-diagnostic-results.json. Re-run the full seven-arm diagnostic without RGB_FOCUS_ONLY; use RGB_FOCUS_ONLY=1 for the three focused combined-image probes.