PRR-1: Provenance Retention Rate — Benchmark Specification (Draft Working Specification v0.1)

PRR-1 proposes a draft benchmark protocol for measuring provenance retention across inference chains in generative reasoning systems. The specification operationalises a central claim of the companion framework paper, Provenance Collapse: Why Hallucination Is a Metadata Failure, Not Merely a Generation Failure (Zenodo, 2026), namely that generative systems fail not only by producing incorrect outputs, but by losing track of the epistemic status of claims as reasoning depth increases. The benchmark introduces a six-state provenance ontology (VERIFIED, INFERRED, SPECULATIVE, ANALOGICAL, WITHDRAWN, UNKNOWN), explicit state-transition rules, perturbation methodologies, annotator agreement protocols, and two primary metrics: Provenance Retention Rate (PRR) and Provenance Validity Accuracy (PVA). The specification also defines invalidation conditions under which the benchmark itself should be considered insufficient. PRR-1 is not an empirical benchmark result and does not claim validation. Annotation reliability, ontology completeness, metric sensitivity, realistic-task transfer, and provenance coherence under revision remain open problems requiring external review and implementation testing. The document is intended as a technical-note companion to the broader Provenance Collapse framework and as a reproducible starting point for experimental evaluation of provenance-aware reasoning architectures. Keywords: provenancehallucinationbenchmarkgenerative AIreasoning systemsepistemic trackingmetadata integrityLLM evaluationannotation reliabilityinference chainsprovenance collapsebenchmark protocolAI safetyreasoning benchmarksepistemic provenance