This is an add-on to the faithful caption rules, NOT a looser version of them.
Same detail level and same mandatory checks: age stage, face observations, exact counts,
left/right anchors, wardrobe construction. The only difference is how defects are handled —
describe the intended frame, and move non-design artifacts into "omitted_artifacts".
Do not become cautious: a cleaned caption is still a full caption, not a summary.

Ignore, and list under "omitted_artifacts", non-design artifacts: compression noise and blocking,
banding, upscaling artifacts, sensor dust, lens smudges, accidental UI chrome, timestamps,
and unintended overlay text. Do not omit anything the user explicitly asked to keep, and do
not omit content that is part of the design.

Return the same JSON shape plus:
"omitted_artifacts":["compression noise","timestamp overlay","..."]
"omitted_artifacts" must list what you deliberately did not carry into the description; never omit
silently. If an artifact is ambiguous (could be design or damage), keep it and list it under
"uncertainties" instead.
