Open-source agent tracing with evaluation templates, from Arize. Here is what ProofAgent adds on top, and why most teams end up running both.
Phoenix is built for inspection after the fact. Nothing in it generates adversarial pressure, so it cannot tell you whether an agent holds when a user claims authority it does not have or a document carries an injected instruction. It produces no compliance artifact.
Governance, not just evaluation. A score is not a decision. ProofAgent carries the agent registry, the AI-BOM, the owner of record and the audit trail alongside the evaluation, so "who signed this off, against what evidence, and what changed since" has an answer. Evaluation tools produce numbers; governance produces the record an auditor asks for.
Policy as code, enforced in CI. Release rules are written as code and evaluated in the pipeline, so a failing threshold blocks a build rather than filing a ticket. A policy in a document is a suggestion; the same policy in CI is a control.
PAI, one readiness number. The ProofAgent Governance Index collapses behaviour, context, compliance and governance into a single production-readiness score computed from a real evaluation run, so a release conversation has one number to argue about instead of a dashboard nobody agrees how to weigh.
Human-on-the-Bridge, published methodology. Expert judgment is curated once, upstream, into reusable adversarial traps, juror personas and scoring rubrics, and small evaluator models then stress-test frontier-grade agents at scale. Published as arXiv:2606.16871 alongside four further papers on adversarial testing, context engineering, memory control and the readiness index — so the methodology can be read, criticised and reproduced rather than taken on trust.
Stress the agent across hundreds of adversarial probes in ProofAgent, then open the failed runs in Phoenix to debug them span by span.
The ProofAgent Harness is Apache 2.0 and self-hosted, with your own model keys, so it is free to run at any scale. It is not a drop-in replacement though: open-source agent tracing with evaluation templates, from arize is what Arize Phoenix is built for, and ProofAgent is built to decide whether an agent should be released. Teams commonly run both.
Phoenix is built for inspection after the fact. Nothing in it generates adversarial pressure, so it cannot tell you whether an agent holds when a user claims authority it does not have or a document carries an injected instruction. It produces no compliance artifact.
Stress the agent across hundreds of adversarial probes in ProofAgent, then open the failed runs in Phoenix to debug them span by span.
No. Arize Phoenix is a developer tool and does not claim to map evaluation evidence onto regulatory frameworks. ProofAgent assesses compliance across 25 of them and produces a signed readiness report, which is the artifact an auditor asks for rather than a dashboard screenshot.
The ProofAgent Harness is Apache 2.0 and runs on your own infrastructure with your own model keys, so you can score an agent before talking to anyone. Read the governance framework free in AI Agent Governance, or see the full four-way comparison.