We introduce SafeRoute-MAS, a coverage‑aware, risk‑calibrated procedure that, using only pre‑integration information, either retains all specialist messages or discards a single one.\n\nOn a sealed cache of 5,328 committees, the source‑compatible operating point gains +1.80 percentage points (95% CI [+1.13, +2.52]) while removing 12.76% of message tokens.\n\nWe also run 4,580 controlled integrations over a stratified sample of 458 committees on a public gpt‑oss‑120b endpoint. The sample estimate is -0.09 points; design weighting to the original workload yields +0.94 points (95% CI [-1.31, +3.50]) and an 18.72% reduction in acted‑input tokens.\n\nA weighted token‑matched comparison likewise does not establish a selector advantage (+1.68 points, CI [-1.27, +5.10]).\n\nOnly 13 of 88 cached non‑neutral labels (14.8%) retain direction by repeated‑call majority. Expected‑corruption UCL is 4.60% after weighting, but only 2 of 11 prespecified marginals pass the support/risk rule.\n\nWe therefore propose a numerical qualification with a Keep‑All fallback, applied retrospectively here rather than validated prospectively.\n\nReview: The approach demonstrates notable token reduction while preserving specialist messages, yet statistical significance and risk guarantees remain limited.