Back to projects

AI Reasoning Architecture

ControlArc

A reasoning architecture for separating structure from expression in AI-assisted work.

Problem space

Most AI workflows collapse source, interpretation, structure, voice, formatting, and final expression into one generated response. When that happens, it becomes hard to inspect what the system understood, what it inferred, what it preserved, and where judgment entered.

Project response

ControlArc explores a different pattern: externalize the durable structure first, then let expression happen as a later layer. The system is organized around source, structure, approval, output, trace, and evaluation.

Current shape

The current project is a public method/tool site and early CLI-shaped proof. It is less a finished product than a working architecture for thinking about inspectable AI systems.

Design decisions

  • Separate structural reasoning from expressive output so review can happen before prose hides the logic.
  • Treat source boundaries, approvals, and trace as first-class parts of AI workflow design.
  • Make structure reusable so the same underlying intelligence can support multiple expressions.
  • Keep human judgment visible at the points where intent, constraint, and output need review.

WIP status

  • A concrete example run that shows source material becoming traceable output.
  • More explicit buyer language for teams that need governance but do not know GCA.
  • A sharper boundary between public method, tool, and consulting capability.

How it connects

ControlArc points to a core belief behind editing.help: before AI systems express ideas on our behalf, we need better ways to inspect and approve the structures those expressions come from.