Knowledge Base Authorship
Every Knowledge Base page and document records who wrote it — a person, an AI step, or an AI draft a person reviewed. How the label is assigned, where it appears, and why you cannot set it yourself.
Every page and document in a Knowledge Base carries an authorship label recording who wrote its content. The label is assigned by Scrydon at the moment the content is written and shown on the entry's Metadata tab in the Knowledge Base explorer.
Today the label is a record. It is written, stored and shown; nothing yet reads it to change which sources reach a model. Ranking on it is a separate change and this page will say so when it ships.
The three labels
| Label | Meaning |
|---|---|
| Human generated | Authored by a person. No AI run contributed the text. |
| AI generated | Produced by an AI step and unchanged since. |
| AI generated · human reviewed | Produced by an AI step, then edited by a person. |
An entry written before this feature shipped may carry no label at all. That means unknown, not "human" — see Unlabelled content below.
A transcript is Human generated, not AI generated. A machine transcribed the audio, but the words are what a person said. The same applies to session notes. Only an agent's own analysis of a recording — a summary, an assessment, an enrichment — is AI generated.
You cannot set the label yourself
Authorship is derived by the server from who is writing, never accepted as a field on an API request.
This is deliberate. The label is intended to become an input to retrieval, which will reserve part of every result window for human-backed content. That makes "Human generated" worth something — so a caller able to declare its own output human would be able to buy itself a reserved slot and defeat the preference the label exists to enable. It is derived from the credential you authenticated with, which you cannot choose per request.
There is no UI control and no API parameter to override it.
How the label is assigned
In a process flow, each writer is a distinct step with a known author, so the label follows the step:
- an operator's upload, a meeting transcript, session notes, a map decision → Human generated
- an AI document action → AI generated, or AI generated · human reviewed if you edited the draft before it was completed
- an agent's assessment, summary, or enrichment page → AI generated
Editing an AI-generated page by hand moves it to AI generated · human reviewed.
It can move back, and the reason matters: if a later agent run rewrites that page's body, the text you reviewed no longer exists, so the page is re-labelled for whoever wrote the new content. Keeping the reviewed label would claim a review of text nobody has read. Your review survives every change except one that replaces what you reviewed.
Through the API or the SDK, the label follows the credential:
- a write performed by a workflow step — where the Knowledge block's Create Document operation runs → AI generated
- a write from an interactive browser session → Human generated
- everything else — API key, OAuth mini-app, delegated background execution → no label, i.e. unknown
An API key does not tell Scrydon whether a person or an agent is on the other end, so it records neither. That is not a gap to be closed by guessing: labelling such a write AI generated would deny a real operator's document its reserved slot, which is the exact problem this feature was built to fix, while labelling it Human generated would hand that slot to content nobody can vouch for. Unknown does neither, and it is the honest answer.
Where you see it
- Knowledge Base explorer — on the entry's Metadata tab, beside its classification. Entries with no label read Unknown.
- Process flow task output — the same three labels on an action's output card.
It is deliberately not shown on every row of the file list. Provenance is a property of a document, so it belongs with the document's other properties; repeating it per row crowded out the filename in the narrow workspace pane and cost every row a render, including the majority that have nothing to show.
What the label is for
The label is a durable record of provenance, and it is what will let retrieval tell an uploaded source apart from the summaries and restatements that grow around it. Recording it is the prerequisite for that preference.
It is also a transparency record in its own right: it answers "was this written by a person" for any entry in the Knowledge Base, at the time it was written, rather than re-deriving an answer that could change later.
Unlabelled content
Content written before this feature carries no label, which Scrydon treats as unknown.
Unknown is not the same as human. An unlabelled entry keeps exactly the position relevance earned it — it is neither promoted nor demoted. Your existing Knowledge Bases therefore behave as they did before, and labels start appearing on newly written content immediately.
An operator can label historical entries in bulk during a maintenance window. That is optional, and it only labels what it can evidence — pages an AI step demonstrably wrote, identified by where they sit in a process flow's Knowledge Base. Everything it cannot evidence stays unknown rather than being declared human, because a bulk sweep that guessed would put the guess on record as provenance.
Related
- Knowledge Base Clearance — who is allowed to see an entry, which is applied before authorship is considered
- Knowledge Base Ingestion — how content gets into a Knowledge Base
- Knowledge Editing — editing a page, and the review transition