HCMA Paper 5: Composable Temporal Governance for Semantic Retrievers

HCMA Paper 5: temporal governance (version collapse at consolidation time) as a composable, retriever-agnostic module. We show that applying governance to any retriever closes the temporal version-drift gap without retraining and without sacrificing paraphrase robustness — semantic matching and temporal validity are orthogonal, composable concerns. Key results (Natural Questions, plausible temporal drift): • Hybrid (pretrained MiniLM + governance): 0.993 paraphrase / 1.000 verbatim / conformal coverage 1.000 (α=0.05), vs pretrained-ungoverned 0.447 / 0.432 / 0.004. • Composability (governance lifts paraphrase accuracy): BM25 0.092→0.962 (+0.870), MiniLM 0.440→0.989 (+0.550), MPNet 0.229→0.989 (+0.760). License: CC-BY 4.0 (paper/docs/data); Apache-2.0 (code, SPDX headers inline). SHA-256 integrity manifest: manifest.json