192What I say, what I think

Generate this image now. Draw it directly - do not describe it, explain it, or ask questions.
Output one landscape A3 poster, 420 x 297 mm, aspect ratio 1.414:1. Use the largest supported resolution. Final print target: 4961x3508 pixels at 300dpi. Keep all text at least 10 mm inside the edges.

[Purpose]
Create one two-column social meme for a single everyday situation, not an app screenshot or comic sequence.
[Scene]
Read "【Situation】". Spoken words: "【What I say】". Private thought: "【What I think】". If either equals "Choose for me", invent a brief fictional caption of up to six words for that side. Otherwise preserve the exact supplied wording. A generated caption is not a documented quote. Do not attack a named person.
[Attached photo]
No photo and no human character. One clean speech bubble on the left, one scratchy pencil thought bubble with two connected circles on the right.
[Text layout]
Equal ice-blue left and muted-coral right halves with a charcoal vertical divider. Exact headings "I SAY" and "I THINK" at y25-31%. Large corresponding messages inside the two bubbles at y40-59%,at most three short lines per half. Exact situation appears once at y74-79% across both halves, without a box. Top22% and bottom18% remain empty. No generic title.
[Style]
Fixed matte editorial graphic, graphite pencil handwriting, off-white bubble interiors and subtle paper grain. No portraits, emoji collage or gauges.
[Visual hierarchy]
Both replies equally large; labels secondary; situation one smaller readable line. Words stay within each half.
[Lighting and texture]
Flat ink and paper, no3D. Clean versus scratchy bubble contours provide the visual contrast.

[Text rules]
Use only the requested visible copy in the target language. Resolve an automatic-copy instruction into an original short line; never print the instruction itself. Preserve user-supplied names, spelling, numbers and punctuation. Input labels, section headings and placeholders are not visible copy. Every optional field stays on its own instruction line so that leaving it empty removes only that item. Never add a date, price, discount, quantity, prize, origin, business condition, scientific statistic or claim about a real person that the user did not provide. Personal inputs affect content only; keep this one art medium, one layout and one finished result type.

[Avoid and delivery]
Return exactly ONE complete raster image directly. All requested elements belong on the same canvas. Do not run code, provide an explanation instead of an image, create separate panels as separate files, require another message, offer variants, or ask the user to cut and assemble anything. No device mockup, promotional wrapper, brand logo, packaging imitation, existing fictional character, artist signature, celebrity or watermark. Finish the selected canvas edge to edge. Maintain the fixed composition when the selected aspect ratio changes by adapting spacing, without cutting important parts or adding a second result.

[Legibility]
Keep essential lettering inside x7-93% and y14-84% of the image. The upper12% and lower15% contain no important text. At a1024x1536 cover, all essential ink is below y215 and above y1290. Use clearly connected natural letters, comfortable spacing and strong contrast. Do not let faces, decorative marks, food highlights or shadows obscure the words. Large lettering must still identify this result in a three-column list. Handwritten copy is legible pencil handwriting, never tiny ornamental cursive.