Research · v1.3.0 · Aug 2026
Derivation log record, an open schema
Eighteen fields, a permissive licence, a changelog, and an invitation to implement it.
Download
The paper, as published
WP-2026-03, The derivation record: an open schema for per-response traceability in conversational AI, version 1.0, 8 pages, CC BY 4.0. The record walked field by field at log v1.3.0, the integrity-tagged state envelope, the Article 12 field-level mapping, the draft normative text, and a worked record from the development build.
The field that matters
binds[]
Most of the record is unremarkable: identifiers, flags, a hash, a result. One field is not. binds[], named bridges_licensed in earlier schema versions, records which relation licensed each reference binding, naming the referent and the relation.
No statistical architecture can populate that field, because no statistical architecture holds a relation it could name. This is the strongest single paragraph we can write about the difference, and it is a statement about record-keeping rather than about quality.
Why it is open
A schema only you use is a file format
The schema is published under a permissive licence with a changelog, at a stable versioned URL, with an explicit invitation to other vendors to implement it. If it is adopted it becomes a standards contribution. If it is not, it is a file format with good documentation.
This costs us nothing we were keeping and it is the sort of artefact that gets cited by people writing about AI transparency.
What the record cannot show
A complete, well-formed record proves that production and comprehension agreed within the grammar. It does not prove that a human reader recovered the intended meaning. The verification is sound with respect to the grammar, not with respect to English.