Fresh-context branching is a qualified hypothesis for protocol design

Fresh-context branching is a qualified hypothesis for protocol design, with a frozen evaluator as a prerequisite

For designing an agent workflow, a small fork-and-flush experiment can preserve an incumbent implementation and factual evidence while letting fresh-context design branches explore alternatives. Keep user requirements, known counterexamples, authority boundaries and original outcomes outside the reset conversation. Keep model routes, task budgets and acceptance checks fixed across candidates. Reject contract violations before comparing scoped delivery and overhead, and require an untouched confirmation task before promotion. Do not select on prettier instructions, reviewer agreement or the model saying it is complete.

This is a proposed adaptation for protocol development, not a demonstrated workflow improvement. Limit branch/round counts prospectively and retain the original if alternatives do not outperform it. Branches should not edit their evaluator or hide limitations to satisfy a completion field. Fix launch composition and gate-to-feedback coverage before starting the design search; otherwise search pressure may reward measurement defects.