Provenance
The record of where information came from, how it changed, and which activity or actor produced it.
Updated 15 September 2026
In practice
A briefing's claim links to the source page, the retrieval date, the version of the document that was read, and the tool call that fetched it.
The distinction
A citation is a user interface for provenance, not provenance itself. A link can point at a source that is outdated, ineligible for this customer, or quoted well beyond what it actually supports.
What an agent job's provenance record holds
job and run ID
source URI or system record ID
source version, owner and effective date
retrieval query and eligibility filters
retrieved passage or record identifier
parser, index, embedding and ranking versions
model and runtime version
tool calls and transformations
output or action influenced
approval and acceptance record
correction, revocation or deletion event
Not every reader needs every field. The system needs enough to investigate an error, reproduce a material decision where possible, honour a deletion request through its derived copies, and tell primary evidence apart from generated interpretation.
Why it becomes commercial
When software produces work a customer pays for, provenance is what allows the supplier to defend the result, correct it, or accept that it was wrong.
A briefing whose claims cannot be traced to a permitted, current source is difficult to review, difficult to correct and difficult to stand behind in a dispute. The same record that makes an agent inspectable is the record that makes the service sellable.
Deletion has to reach the copies
Information flows into chunks, embeddings, indexes, memories, traces and backups. A deletion that removes the original but leaves a derived copy influencing new jobs has not been honoured.
Tracking lineage forward is what makes that reachable, which is why provenance and retention policy are the same design problem.
Sources & context
W3C · 2013
Formalises provenance as entities, activities and agents — the underlying model for describing how a record came to exist.
Source checked 2026-09-13
OpenAI · 2026
Documents search results carrying chunks, similarity scores and the originating file, which is the minimum provenance a retrieval step can return.
Source checked 2026-09-13