Perspectra: A Layered, Provenance-Preserving Framework for Multi-Source News Event Synthesis

This paper reframes multi-source news understanding as an event-synthesis problem rather than an aggregation or summarization task, treating the real-world event — not the individual article — as the basic unit of processing. We introduce a two-pipeline, seven-layer architecture: a continuous Global Discovery pipeline that forms provenance-preserving event records, and a sandboxed Active Enrichment pipeline that safely expands thin or stale events without cross-event contamination. Claim extraction, convergence estimation, and perspective analysis are gated behind sufficient multi-source evidence, and ambiguous cases are resolved through human adjudication rather than automatic thresholds alone.

We evaluate the system against an independent, blind reference clustering covering the entire 4,491-article corpus, rather than a curated subset. Perspectra achieves a pairwise precision of 0.7277, recall of 0.5602, and F1 of 0.6331, while maintaining zero article- or source-count integrity violations across all recovery and maintenance operations — and we report the recall limitations candidly as a negative finding rather than omitting them.