# Edges, boundaries and paths

> How Keryn models trust: edges first, trust boundaries where authority changes, and chains of edges an attacker can follow, each with its evidence.

Source: https://keryn.dev/docs/edges/

Trust lives in relationships. A model, a repository, an agent or a person is not safe or dangerous alone: risk appears where one can reach, call, read, change or act for another. Keryn models those edges first and the things they connect second. This section is the manual for that model. [Why the edges](https://keryn.dev/docs/concepts/why-the-edges/) is the short version.

Diagram: user text sends a prompt, across a trust boundary. Every edge has a kind, two ends, an authority and evidence.

## What an edge carries

- **A kind**, one of eight: reaches, calls, reads, writes, sends, grants, depends on, trusts.
- **Two ends**, the source and the target.
- **The authority** under which it exists: who may make it happen.
- **Its evidence**: the file, line, test or review that supports it, or an explicit statement that there is none.

An inventory of components, the AI-BOM, is still there. It is the list of nodes the edges attach to.

## What to read

| You want to know | Read |
|---|---|
| Where authority changes in an AI system, and what supports each crossing | [Trust boundaries](https://keryn.dev/docs/edges/trust-boundaries/) |
| What the eight kinds mean, with a runnable example of several | [Edge kinds](https://keryn.dev/docs/edges/edge-kinds/) |
| How edges combine into paths, chains and graphs, and where to put a guard | [Paths, chains and chokepoints](https://keryn.dev/docs/edges/paths-and-chains/) |
| What a model may and may not do when it reviews findings | [Agentic triage](https://keryn.dev/docs/edges/agentic-triage/) |
| What Keryn will not do, stated as guarantees | [Guarantees](https://keryn.dev/docs/edges/guarantees/) |
| How to draw your own system and analyse it | [Model your agent on the Trust Canvas](https://keryn.dev/docs/guides/trust-canvas/) |

## The model in one picture

Chain: Untrusted input sends (across a trust boundary), Agent calls (chokepoint), Tool writes (across a trust boundary), Database. A chain. Mint cuts mark boundaries. The heavy edge is the chokepoint: guard it and the chain breaks.

A boundary is one edge. A chain is edges in order. A chokepoint is the element many chains share. Keryn keeps the evidence on every edge, so a chain can be proven, challenged or broken.

## Where you see edges today

The edge model is the direction of the product, and the surfaces show it unevenly. Today:

- **Rules** report a source-to-sink path inside a file and across files, with the proof level.
- **Rule metadata** names the edge and the boundary class each rule inspects.
- **Keryn Web** lists the routes that reach a model and the tools each can call, and exports the evidence graph.
- **The workflow analyser** reports risky combinations in agent exports.
- **The Trust Canvas** holds the full model: boundaries, reach, chains, chokepoints.

There is not yet one model of every edge shared by all of these, and cross-file reach is bounded. See [what Keryn does not do](https://keryn.dev/docs/concepts/limits/).
