Self-Refinement Loop or Tree Search for a Contract-Clause Rewriter?

A feature rewrites a flagged contract clause to remove a specific legal risk (e.g. an overly broad indemnification scope) while preserving the clause's original intent. Early testing with a single CoT call succeeds about 60% of the time (a lawyer reviewer agrees the rewrite both fixes the risk and preserves intent); the most common failure is a rewrite that fixes the risk but drifts from the original intent in a way a second look would catch.

1. Would a sequential self-refinement loop (critique, then revise) plausibly help here? What property of the failure mode makes you say yes or no?
2. Would tree-of-thought search plausibly help more, less, or about the same? Justify by naming the specific structural property ToT needs that this task does or doesn't have.
3. Propose the technique you'd actually try first, and the concrete measurement that would tell you if it's working.