ImpossibleMA: Bounded Synthesis of Cochrane Meta-Analyses Declared Unpoolable

Abstract

Cochrane meta-analyses whose prediction intervals cross the null are conventionally deemed unpoolable, leaving guideline panels with narrative synthesis. Screening 6,229 MetaAudit meta-analyses (68,519 audit rows) for any CRITICAL flag, only prediction_gap fires CRITICAL, so the contribution is the envelope construction, not multi-module triage. The Possibility Envelope—(lower, upper, point|None, min_info, assumptions)—is computed by an adversarial engine enumerating inclusion-rule combinations and reporting REML+HKSJ extremes. 444 of 6,229 (7.13%) carried a CRITICAL flag, all converted to bounded adversarial envelopes, zero unclassified. REML matches metafor::rma below 1e-6 (our house ideal; τ≈1e-7 except the degenerate τ²=0 boundary); monotonicity and full-data containment property-tested across 23 inputs, 182 tests gating every commit. The envelope bounds defensible pooled effects under stated rules, and one named datum—a pre-registered protocol—collapses it to a point; since the bounded set is defined by the same prediction_gap criterion, "unpoolable" is definitional, not external. v1 covers adversarial, k=1, and missing-SE cases; figure-extraction, incommensurable outcomes, and per-MA runs are deferred.