Derivation Links

Status: Design landing Reference epics: INK-827

What a derivation link is

When the World Agent produces inferred content (origin: agent-produced) from other workspace content, the derivation is recorded as a derivation link from the new content to its sources.

A derivation link is a directed relationship:

derived_content -[derived_from]-> source_content

Many-to-many: a derived page may have many derivation links (one per supporting source); a source may have many derivations pointing at it.

Closed to the workspace

Derivation links are entirely internal to the workspace. Sources are always workspace entities — either stored directly or referenced by their workspace identifier. Cross-workspace derivation is not supported; each workspace is a closed world.

Practical consequence: a derivation link cannot point at an external URL, an imported reference document’s original source, or an external model’s reasoning trace. External provenance, where it matters, lives in origin metadata (origin: imported carries the external source citation the import preserved). Derivation links describe intra-workspace support.

First-class relationship

Derivation links are first-class relationships in the workspace graph, not comments or embedded metadata. They are:

• Examinable — visible on the derived content as an affordance, expandable into the full source list.
• Queryable — the graph can be asked “what derives from X?” and “what does Y derive from?”
• Watchable — the retroactive-revision queue subscribes to changes on source entities and flags derived content when sources change.

Storage shape lives in the database schema (see Database Schema) as a derivation_link table with derived_id, source_id, a link type discriminator, and creation metadata.

Three purposes

1. Honest attribution

When the agent later cites inferred content, the derivation lets it trace back to its supporting sources and show them. The World Agent’s property of explicit epistemic standing (see World Agent §3) depends on derivation links: “Here is what I concluded, and here are the 35 pages I concluded it from.”

2. Retroactive revision

When sources change, derived content that depended on them can be flagged. This is the substrate the retroactive-revision queue runs over. Without derivation links there is no graph to walk; with them, the world model can surface “a page this was derived from has changed” to both the user and the agent.

3. Explainability

The user can ask the agent “how did you arrive at this?” and the derivation link is the first layer of answer. Beyond the links themselves, the agent can pull the source pages into context and reconstruct the reasoning; the derivation link is what scopes that reconstruction to a specific, attributable set.

Derivation is not provenance renamed

Derivation links and origin are distinct:

	Origin (provenance)	Derivation (relationship)
Shape	Flat label on content	Graph edge between content
Answers	Where did this enter the workspace from?	What other workspace entities support this?
Cardinality	One value per content item	Many per content item
Decays?	Never (terminal)	Removed only if the derivation is rejected
Fed into	Agent’s kind-of-claim reasoning	Standing computation, retroactive revision, explainability

Inferred content can have many derivations; authored content typically has none — though a user-authored page citing other pages is a natural place for the user to record their own derivations as wiki-links that also generate derivation links, since the act of citing is recording a derivation.

When derivation links are created

Derivation links are attached at the submit boundary. At registration, any tool that crosses the boundary declares whether it is derivation-producing and how derivation sources are supplied:

• Explicit argument — the tool takes a list of source entity IDs as an argument (this is the common case for agent-inference tools).
• Inferred from the agent’s reasoning trace — the submit-boundary adapter reads the agent’s working context for the source entities consulted during the tool call (for tools where explicit source-passing is impractical).
• Not applicable — the tool does not produce derivations (most editing tools; creating a blank page; importing content — imports carry origin, not derivation).

Tool authors declare this once, at registration. The adapter constructs the derivation-link records as part of the world-write it builds (see Submit Boundary §adapter).

Standing is computed over derivation links

Standing is the derived signal over the derivation subgraph. A source’s standing reflects how many canonical derivations point at it. The shape:

• Only canonical content confers standing on its sources. A candidate derivation does not count until it is elevated.
• A source’s standing is computed over the set of canonical descendants, not just direct derivations. If A ← B ← C, and B and C are canonical, A has standing from both.
• Standing accrues; it is not author-set. The author cannot pin standing directly; they can influence it by canonicalizing or retiring derivations.

Storage-wise, standing is computed on demand (with caching as an optimization in the infrastructure layer); it is not stored on content.

Lifecycle of a derivation link

1. Created at the submit boundary when derived content is submitted.
2. Observed by the retroactive-revision queue when its source changes.
3. Modified if the derived content is edited in a way that drops a source (the link is removed) or adds one (a new link is created — but this requires another submit-boundary transit).
4. Retained through the derived content’s lifecycle transitions. A derivation link on candidate content remains when the content is promoted to canonical; the link now contributes to the source’s standing.
5. Retained when the derived content is retired. Historical derivations are preserved; retired content’s derivations no longer contribute to standing accrual on their sources, but the record is not deleted.

Worked example

Moment 2 — Agent derivation shows 35 derivation links being created when the World Agent produces Binding — rules, costs, and edge cases from 35 source pages. Moment 3 — Source edit shows those derivation links in action as the substrate for the retroactive-revision queue: when The Nature of Binding is edited, the graph walks its inbound derivation edges and flags every dependent page.

What this page does not do

• Does not describe the storage schema. See Database Schema.
• Does not describe the retroactive-revision queue that consumes derivation links. See Retroactive Revision.
• Does not describe the standing-computation algorithm in detail. See Provenance §standing.
• Does not describe how the agent reconstructs reasoning from derivations. See Conversation System for the reasoning-trace surface.