Hashing Is Not Retrieval: A Three-Arm Study of Context Accumulation in Small-Model Code Cartography

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Hashing Is Not Retrieval is a prospectively frozen three-arm evaluation of context accumulation in small-model multi-agent code cartography. It compares three ways for an orchestrator to hand work to a worker: 1. FULL HISTORY — retransmit prior worker outputs. 2. OPAQUE REF — replace the previous output with an unretrievable SHA-256 digest. 3. RETRIEVED — resolve the digest through a tenant-scoped content-addressed store and inject the recovered bytes. All arms receive identical source-derived evidence. The apparatus measures the exact message content sent to workers and records token counts, model and provider identifiers, generation IDs, usage, cost, retries, and retrieval proofs. A five-task gate preceded a fixed 50-task controlled benchmark, followed by a separate state-required evaluation across 24 deterministic fixtures and a five-scale Medusa case study. Key result: in the controlled benchmark, RETRIEVED reduced paired peak context by 38.3% relative to FULL HISTORY. The state-required follow-up shows that opaque references fail when state must be recovered, while verified retrieval preserves substantially more structural information. This is bounded-context evidence for automatic orchestrator-resolved state passing inside a specific two-worker code-cartography harness — not a universal claim about autonomous memory, long-context capability, or reasoning quality. The release includes the corrected author-identified preprint, executable apparatus, frozen results, task fixtures, telemetry, and verification instructions. Submission status: This preprint was submitted to the NeurIPS 2026 Workshop on Long Context Foundation Models (LCFM). LCFM is a non-archival workshop; this Zenodo record is a public preprint and does not indicate acceptance or publication. Workshop information: https://longcontextfm.github.io/ Research website: https://abdulmoizahmed.com/research Latest version DOI: https://doi.org/10.5281/zenodo.22848077 Concept DOI (all versions): https://doi.org/10.5281/zenodo.22844754 License: CC BY 4.0.