Coverage Before Compression: Online Call Scheduling for Repeated Long-Context Review

Repeated long-context review amortizes candidate proposal while incurring target-model verification cost on every case, making call scheduling a first-order online design variable. We introduce policy-isolated verification, a paired estimand that holds the output, span partition, candidate order, target model, and call cap fixed, together with EvidenceCall, a coalition-state policy organized by coverage before compression. EvidenceCall tests necessity and direction, continues from verified coalitions, and delays redundancy removal until evidence coverage is secured. At an identical $B=50$ cap over $3,600$ RULER-128K, MuSiQue, and QuALITY examples, it improves decisive-span P@5 by $+9.3$ over Greedy Minimal-Set, with $80.6$ of the margin retained under a literal ranking freeze. Across three frozen proposers, it also leads outside-family MCTS-Subset by $+1.2$ to $+1.8$ P@5, and every paired-bootstrap lower bound is positive. The identified gain transfers to decisions: reviewer yield rises by $31.7$ in a $32$-reviewer mechanism-isolation crossover and by $28.1$ in a $24$-developer code deployment, while a disjoint $36$-reviewer citation deployment improves accepted findings from $3.88$ to $4.28$ per hour over MCTS-Subset. Blind full-context adjudication, external proposer supervision, three held-out benchmarks, and two additional target families preserve the effect. The results establish coverage-before-compression scheduling as a consequential systems principle for repeated review.