Skip to content

Platform Overview

Application Main Screen

Applications

Applications are containers that group together cooperative guardrails and shared data assets. You can think of them as projects. You must create an application before you can create your first guardrail.

An application corresponds loosely to an external LLM-driven application, and is a container for the verification assets for it.

Guardrails

Guardrails are the primary verification assets created in this platform. Guardrails are specific procedures that verify the outputs of an LLM-driven application using a particular technique or along a specific dimension. Guardrails come in in three primary categories:

  1. Agentic - General purpose guardrails that utilize an LLM for "reasoning" about the output of another LLM. These have the benefit of generality, and can address a wide swath of hallucination and other error classes, but ultimately suffer from using an LLM to "reason". They can effectively reduce error rates, but cannot provably eliminate them.
  2. Neurosymbolic - Integration of the LLM outputs into formal symbolic logic can provably perform reasoning tasks with perfect accuracy (subject to garbage in, garbage out). The tradeoff is that the symbolic reasoning must typically be constrained to a specific policy. Regarding the bridge between natural language and the symbolic language, Jaxon's approach to neurosymbolic guardrails allows LLMs to perform data extraction instead of verification reasoning (an easier and better-suited task) for automation, and optional human review of extracted data for guaranteed consistency.
  3. Manual - Some processes warrant direct human verification. Jaxon offers a manual Human Review guardrail that can implement these manual steps.

Human Review

Jaxon offers a Human Review dashboard for manual verification and correction of guardrail inputs and outputs. This can be used for guardrail testing, but is intended for use by analysts/reviewers who are incorporated into some verification processes. See above and elsewhere in this documentation for specific usages of the Human Review dashboard.

Dashboards

The system dashboards give an overall perspective on guardrail usage and statistics.

Logs

System logs are available both in Docker (programmatially) and here for manual review and monitoring.