Forge Graph

The Forge Graph is your AI organization as a connected system: every engineering artifact and the real relationships between them.

Why a graph

AI systems are graphs, not tables. A prompt is used by an agent, which is measured by an evaluation, which uses a dataset, which supports a decision, which produces knowledge that bears on three other artifacts.

Every interesting engineering question is a traversal:

• What would break if I changed this dataset? — walk the dependents.
• What does this evaluation actually cover? — walk the dependencies.
• Where is this prompt reused? — walk the edges.
• What is this failure connected to? — walk the neighbourhood.

None of these are answerable in a list view, which is why most tools cannot answer them at all.

            ┌───────────┐         ┌───────────┐
            │  Provider │         │  Dataset  │
            └─────┬─────┘         └─────┬─────┘
                  │ serves              │ measures
                  ▼                     ▼
   ┌────────┐  ┌───────┐         ┌────────────┐
   │ Prompt │─►│ Agent │────────►│ Evaluation │
   └────────┘  └───────┘ measured└──────┬─────┘
      used by              by           │ produced
                                        ▼
                                 ┌─────────────┐
                                 │ OBSERVATION │
                                 └──────┬──────┘
                                        ▼
                        DECISION ──► KNOWLEDGE

The reasoning overlay

Toggle Show reasoning and the derived objects — observations, claims, decisions, evidence, knowledge — appear as nodes attached to the artifacts they came from.

This is the part with no equivalent elsewhere: you can literally see the thinking layered over the system, rather than having it buried in pages. Selecting a knowledge node opens its own page.

Colour carries one meaning

A rule inherited from the design language and worth stating, because it is what keeps the graph readable: structural hues say what a thing is; the verdict palette says how it is going. The two are never mixed. A node's shape and colour tell you it is a prompt; its state tells you whether it is healthy, needs attention, is at risk, has failed, or is simply not yet known.

Focus

Click any node to focus its neighbourhood. Arriving from elsewhere — an artifact page, a Brok answer, an investigation — focuses the graph on the node the answer was about, via /knowledge?focus=<node-id>.

Node ids are stable and composite: prompt:<uuid>, evaluation:<uuid>, decision:prompt-version:<uuid>. The same id identifies the node everywhere in the product, which is why a Brok answer and a graph selection stay in step.

Where it appears

• `/knowledge` — the full graph.
• Beside Brok — a compact graph in the workspace rail, following whatever the conversation is

about.

• Inside an investigation — the same compact graph, following the timeline event or record you

select.

• From any artifact — See in graph on the Evolution tab.

Graph versus Execution Graph

Two different graphs, two different jobs — a common point of confusion:

See also: Evolution · Registry · Engineering Intelligence