How the context graph works
Wires are context
In ThoughtDAG, a wire is not decoration. When you ask from a node, incoming structural paths determine which upstream conversations and materials are compiled into the model request.
Evidence A ─┐
├─→ Current conclusion ─→ Next question
Evidence B ─┘After disconnecting Evidence B → Current conclusion, Evidence B remains on the canvas. You can continue exploring from it or reconnect it later, but it has been removed from this path's model context. Questions continued from “Current conclusion” no longer send Evidence B to the model.
Why a DAG
- Directed: context flows from earlier material and reasoning toward later questions.
- Acyclic: a node cannot recursively include itself in its own history.
- Branchable: alternatives can be explored without overwriting the original path.
- Mergeable: selected evidence paths can explicitly meet in one synthesis.
Canvas-visible is not model-visible
| State | Meaning |
|---|---|
| Canvas-visible | You can still inspect and reconnect the content. |
| Model-visible | The content is selected by the current context paths and included in this request. |
This distinction preserves discarded ideas, old assumptions, and failed experiments without forcing them into later answers.
What this does and does not explain
ThoughtDAG can inspect and change what the model received. It cannot prove why the model internally produced a particular answer. Input provenance and internal causal explanation are different problems.
Workflow canvases often use wires for execution order or data dependencies. ThoughtDAG's core wires represent model-visible context. If runtime execution is added, execution state should remain distinguishable from context dependency.
For external agent sessions, see Session Atlas: it creates a read-only source mirror, not an in-place editor for the original log.
From an interaction rule to a testable intervention
“Wires change what the model receives” is not only an interface rule. It can be tested under controlled conditions. ThoughtDAG's Context Intervention Benchmark propagates an erroneous claim through several turns, then compares source-only pruning, contaminated-subgraph removal, and dependency-ordered recomputation. The final question and scoring rule stay fixed; the context state changes.