Abstract

We specify a retrieval-augmented generation (RAG) response format in which every document a retrieval server returns to a language model carries a provenance envelope binding the served text to a cryptographic hash of a named, fetchable source file, at least one timestamp independent of the server (an RFC 3161 time-stamp authority, an OpenTimestamps Bitcoin anchor, or a transparency log), and, where available, a persistent identifier (DOI). The envelope also carries a scope statement declaring what the hash covers relative to the served text (whole source, excerpt, or derived rendering), an explicit "no anchor" statement rather than silence when none exists, licence terms, and a runnable verification instruction (fetch, hash, compare, verify anchor). Target failures: staleness, truncation, paraphrase drift and substitution; staleness is caught by re-fetching the named source. Constraints: provenance is an admission predicate, never a ranking weight; the server does not summarise; the licence gate lives in the served index, not the request path; multiple serving surfaces pin one published artifact so their hashes are identical.

The same envelope, served without payload, is the complete public integrity surface of an append-only log in the Certificate Transparency pattern (RFC 6962/9162): signed tree heads, inclusion and consistency proofs, and a published scoring function frozen per period. Coverage is provable only against receipts held by the parties whose events are logged; split views are detected by automated tree-head comparison among independent observers; entry contents are never queryable. The mechanism proves byte identity and prior existence, not truth or currentness.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS