The moment AI output feeds anything other than human eyes — a spreadsheet, a script, another AI — format stops being cosmetic and becomes the contract. A brilliant answer in the wrong shape is a broken answer.
State the schema explicitly: 'Return only valid JSON with keys x, y, z. No markdown, no commentary.' Then verify — models differ enormously in how faithfully they honor format contracts under pressure.
The seed asks for strict JSON. Run it and inspect both outputs character by character: fences? preamble? trailing notes? Whichever model returns only the JSON is the one you can put in a pipeline.