Move a working prompt to a cheaper model without losing the output

You are helping me port a prompt that already works on an expensive model to a cheaper or smaller one. I will paste the prompt and describe the output I depend on.

Work in this order:
1. WHAT THE PROMPT ACTUALLY RELIES ON — list the capabilities it needs: long-context recall, instruction following over many rules, arithmetic, format discipline, domain knowledge, or reasoning depth. Rank them by how much the output would suffer if the model were weaker at it.
2. LIKELY BREAK POINTS — the two or three instructions a smaller model will drop first. Quote them from my prompt.
3. REWRITE — the ported prompt in full. Make the structure explicit: one instruction per line, numbered output, an example if the format is unusual.
4. EVAL — five test inputs (including one edge case and one adversarial), and for each, the pass criteria that can be checked mechanically.
5. VERDICT CONDITION — the result on those five that means ship it, and the result that means the expensive model is worth the money here.

Do not claim any specific model will pass. Do not add capabilities the original prompt did not have.

PROMPT: {{paste the current prompt verbatim}}
OUTPUT I DEPEND ON: {{what downstream consumes it, and what a failure costs}}
TARGET MODEL: {{model, context limit}}