Provenance Is Not Independence: The False-Corroboration Exclusion Rule for Multi-Agent Memory

Multi-agent memory systems increasingly reconstruct provenance chains to improve trust in shared knowledge. The leading production architecture, MemClaw (arXiv:2606.24535), successfully reconstructs 100% of depth-four derivation chains with correct writer identity. Yet provenance reconstruction alone does not prevent false corroboration: when two agents agree because they share a common epistemic ancestor, a system that counts their agreement as two independent confirming sources is miscounting. This paper names the gap and proposes the False-Corroboration Exclusion Rule: before any corroboration tally, collapse all same-root citations to a single unit of warrant. The rule is derived by analogy from hearsay-within-hearsay doctrine in evidence law (FRE 805), where repeated out-of-court assertions tracing to the same original declarant are recognized as one source multiplied, not many sources agreeing. The Kish effective-sample-size formula provides a computable, continuous form: n_eff = k / (1 + (k-1) * phi_bar), where phi_bar is mean pairwise agreement among sources. Empirical calibration from LLM judge panels (Kohli, 2026) yields n_eff = 2.18 for a nominally nine-member panel, an independence ratio of 24.2%: nine voices, fewer than three effective votes. The W3C PROV-DM standard supplies the substrate for the check as a deterministic set operation over lineage, requiring no model calls in the hot path. We argue that governing provenance and governing independence are separate problems, that current production systems solve the first and skip the second, and that the exclusion rule closes the gap.