Skip to main content
Fella separates model judgment from data execution and saved analytical state. The model can direct an investigation, but the app and Rust engine control what it can inspect and how results are recorded.

The main parts

The renderer communicates through an allowlisted preload bridge. Electron starts the Rust engine as a local JSON-lines sidecar. The model has no direct filesystem path, database handle, shell, or Electron API.

One analytical request

  1. The user asks in a conversation and may select a mounted workspace.
  2. The harness snapshots the workspace revision and assembles the immediate request, relevant context, Question scope, and tool schemas.
  3. The selected model directs the analysis. The engine runs only registered operations and returns bounded results.
  4. The harness records each tool result as evidence and runs deterministic checks against the saved work.
  5. The canonical answer is saved with its analysis run. When useful results support a structured view, the model composes a JSON-render document bound to those results.
  6. The renderer validates and displays the view. The saved answer and evidence remain available if presentation composition or rendering fails.

Product data model

An explicit mounted folder is the workspace. It is not the window’s panel layout. A workspace can contain many conversations and Investigations. Creating an Investigation creates its linked conversation; an ordinary new conversation does not automatically create one. The Investigation owns a brief, Questions, source selections, definitions and assumptions, linked analysis runs, and Findings. Each run preserves the workspace revision and context it used. Saved artifacts and reports live separately from the transcript, so deleting a conversation can retain intentionally saved work and its supporting records. The user-facing report destination is Reports. A report document can place exact revisions of workspace-owned artifacts and keep authored narrative. Report placements do not rerun analysis.

Analytical presentations

JSON-render is a presentation layer over engine-produced results, not a second analytics runtime. The model selects registered components and composes an answer-specific hierarchy. Tables and charts bind to accepted result IDs and fields; presentation controls are limited to display state or explicit user-triggered follow-ups. The renderer receives no permission to query files or execute arbitrary code. A structured view can be validated independently of the canonical prose answer. If validation fails, the answer and analysis record remain inspectable.

Ownership map

The split makes the model replaceable without changing local ingestion, tool boundaries, or verification. It also keeps presentations from becoming an alternate source of analytical facts.

Limits

  • A valid computation can still use the wrong interpretation or population.
  • Verification checks recorded work and known failure patterns; it cannot prove every conclusion.
  • Hosted providers receive prompt context and the results requested during analysis.
  • The fixed tool registry has no shell, file-write, web-search, or general connector capability.
  • The report dependency view discovers explicit report placements and checkpoints, not every prose reference or external copy.

The harness

See the control flow from a question to its recorded result.

The analytics engine

See how supported files become bounded local analysis inputs.

Verification

Understand what checks can and cannot establish.

Product model

See the user-facing objects and their relationships.