Model-agnostic Agent Skill: keep multiple hypotheses alive, test them against evidence, revive alternatives, collapse to the best-supported answer. Ships the skill contract, a reference branch controller and a benchmark harness. No model evaluation published yet, so no performance claim.
