Appendix A — External examples

The following table lists examples of external projects that borrow trusted mini-agent principles.

Project	Description	Trust
Clinical Trials Agent	A commons agent for clinical trial reporting. It answers questions about a simulated study’s safety, disposition, exposure, demographics, and laboratory results over synthetic CDISC ADaM datasets. Leverages the NEST TLG Catalog.	Inherits the commons model (see below). The level of trust depends on the commons mode set in the deployed app.
commons	commons is a framework to build custom data analysis agents with pre-loaded deterministic code and datasets. When the user asks for a fitted model or summary statistics, the agent can select these pre-loaded calculations instead of ad hoc AI-generated code. This is similar how the example simulation agent earns trust.	With mode = "trust only", commons >= 0.1.1.9000 only runs trusted code, and it is a full trusted mini-agent following all three rules. With mode = "normal" (default), commons can run untrusted AI-generated code (bending rules 1 and 2), but it informs the user about the level of trust behind every calculation.
retroglyph	An R package that reconstructs individual patient survival data from published Kaplan-Meier curve images. The AI model does very little (reading axis labels, color elements, etc.), and the harness is mostly deterministic. The user assesses the success of the data reconstruction by visually comparing the reconstructed curves to the original image and numerically comparing reconstructed summary statistics to the summary statistics reported in the original publication.	retroglyph is a full trusted mini-agent following all three rules. It implements hard guardrails on the selection and wiring of tools (rules 1 and 2), and the human oversight tasks are visually intuitive and feasible (rule 3).