Vre

VRE — Volute Reasoning Engine

Stability

As of **v1.0.0**, VRE is stable. The **public API** is everything exported from the top-level `vre` package and `vre.core` — the `VRE` class, the `vre_guard` decorator, the repository backends, and the structured result and gap models they return. Anything reached through a private submodule path or a leading-underscore name is internal and may change without notice. The project follows [semantic versioning](https://semver.org/spec/v2.0.0.html): breaking changes to the public API land only in major releases and are preceded by a deprecation period with runtime warnings. Every change is recorded in the [CHANGELOG](CHANGELOG.md).

**Epistemic enforcement for autonomous agents.**

VRE is a Python library that gives an autonomous agent an explicit, inspectable model of what it knows, and refuses to let it act past that boundary. Before a guarded tool executes, VRE checks that the concepts the action touches are grounded, to the depth the graph requires, in an authored knowledge graph. If they are, the action proceeds. If they are not, it is blocked and the specific gap is surfaced as a structured object rather than a generic error.

It is deliberately narrow about what that means. VRE enforces consistency with a *modeled* domain. It does not manufacture understanding, verify that the model is correct, or check the truth of what the agent does. It is **not** a permissions system, not a sandbox, not a correctness guarantee, and not a hallucination filter. What it provides is one thing, done deterministically: an action executes only if it is epistemically justified against the graph and permitted by the policies on the edges it touches.

That last point is the distinctive one. The constraints VRE enforces are not a ruleset bolted alongside the knowledge. They are carried by the graph's own