R / 02 — RESEARCH RECORD
Hash-chained verbatim quotation for verifiable provenance
A preprint about integrity between stored fragments, retrieved quotations, and an audit journal, with its negative results preserved.
01 / BOUNDED RESULT
What happened
The quotation-plus-journal layer detected and localized 100% of store-only tampering in the minimal benchmark.
02 / METHOD
How it was tested
The layer answers with verbatim quotation of stored fragments and records a cryptographic relation between every output and its source material. The minimal benchmark mutates the store after writing and measures detection and localization.
The comparison retains a deliberately simple baseline: echo plus journal. This separates the value of verifiable provenance from any retrieval-engine-specific advantage.
03 / NEGATIVES AND LIMITS
What the result does not support
The result covers store-only tampering in a minimal benchmark. It does not measure retrieval relevance, content truth, or an attacker that also controls the journal or its anchor.
Echo plus journal performed identically to the quotation layer. The result is therefore not a Hyphae-exclusive advantage; a chain-aware adversary requires an external anchor.
The evidence shows that the source-output relation can be made verifiable under this threat model. The Hyphae name in the preprint identifies its experimental lineage, not proof of current public-release behavior.
04 / ARTIFACTS
Read, verify, and cite
Gutierrez, Mario. “Hash-Chained Verbatim Quotation: A Verifiable Provenance Layer for Grounded Retrieval.” preprint v0.2.0, 2026. doi:10.5281/zenodo.20436643.
Back to all research