Technology
Why a human-governed knowledge workflow fits
How durable files, proposal-based AI, revision checks, private audit, preview, and explicit publication make Connectome Studio suitable for consequential knowledge work.
Knowledge needs a durable centre
Connectome Studio keeps Markdown and YAML as canonical source. Cards hold focused relational knowledge; pages compose directories, summaries, and other structural views. Git records change over time, while each website owns its presentation and publication.
This is suitable for personal and organisational knowledge because the core remains readable outside the application. The content is not trapped in one editor, model provider, or database schema. A human can inspect the files, review a diff, and rebuild the site using ordinary tools.
AI proposes; people decide
Model inference is useful for discovering possible links, suggesting a split or merge, drafting text, and explaining evidence. It is also uncertain. Connectome Studio therefore separates model output from editorial truth.
AI work becomes a proposal with evidence, a source revision, and a structured operation. Canonical content changes only after explicit approval and revision checks. Agent conversation stays separate from an editable artifact, so explanation cannot quietly become publication.
The workflow creates trust boundaries
- Search and deterministic analysis narrow the evidence.
- An editor deliberately invokes AI with bounded tools and context.
- The model returns speech, evidence, and an optional proposed artifact.
- The Studio validates identifiers, relationships, schemas, and source revisions.
- A person accepts, revises, or dismisses the proposal.
- Preview builds render the selected project and run its validation.
- Publication requires a fresh production build and a separate explicit decision.
Private conversations, model exchanges, and audit traces stay in project-local ignored storage. They do not leak into public pages or become canonical knowledge.
Why this matters
The approach combines the adaptability of AI with the accountability of editorial systems. It supports cloud and local models without letting either become the source of truth. It also preserves project isolation: content, settings, proposals, histories, audits, previews, and deployments remain scoped to the active project.
The result is slower than unconstrained generation in the moments where judgement matters—and intentionally so. It is designed for knowledge that should remain understandable after the model, interface, or hosting provider changes.
Architecture basis: accepted ADRs 003, 008, 009, 010, 016, and 018. The lifecycle contract in ADR-012 is accepted but not yet implemented, so this page does not claim lifecycle enforcement is complete.