Selective Context Under a Claim-Scoped Evidence Contract: A Synthetic Runtime Characterization of Token-Bleed R5

This technical preprint reports Token-Bleed R5, a synthetic runtime characterization on one named local endpoint. It compares raw full-context stuffing, a bundled compact candidate-context route, and a deliberately name-only lexical negative control across a frozen 20-seed configuration. At zero injected candidate-generator misses, the bundled route used 96.9% to 97.9% fewer mean prompt tokens and achieved higher mean F1 than verbose full context. It did not meet the separately prespecified 3x prompt-token-ratio rule versus the lexical negative control. The candidate generator is oracle-controlled, and the bundled route changes both context membership and representation. These are not production, customer-data, ROI, cross-model, retrieval-performance, or universal economic claims. The deposited files form an auditable derived-evidence package: they retain the frozen contract, privacy-safe derived evidence, paired statistics, and claim-scoped decisions. The original raw report remains private because it contains host-identifying provenance; its SHA-256 commitment is retained only for chain of custody.