A small team working with ML engineers, security leads and compliance leads.
About
Governance work should be done by agents. Governance decisions should not.
FairMind started as a bias-detection toolkit and became a governance platform when we noticed where the real cost was: not in running a test, but in everything around it. Collecting the evidence, mapping it to fifty frameworks, getting a person to look, and proving later that they did.
AGPL-3.0 platform, Apache-2.0 SDK, CLI and MCP server. The product you evaluate is the repository you can read.
Environmental evidence for AI systems, with limitations attached.
Five rules we don't trade away.
They are written into the workflow validator, the gate, and the ledger, not only into this page.
- Deny by default.If an approver is unreachable or no policy covers an action, nothing is recorded or changed.
- Capability truth.Every evaluator and every workflow step says what it does today: validated, external provider, metadata only, unavailable, or manual.
- Evidence over claims.A measured failure outranks an attestation. Readiness is never certification.
- The agent is governed too.Scoped identity per step, a tool allowlist prompts cannot widen, a hash-chained trace, a kill switch per run.
- Gaps shown, not hidden.Framework journeys list every requirement, including the ones nothing covers yet.
Why evaluations and workflows, not a checklist.
A checklist tells you what a framework asks. It cannot tell you whether the agent blocked an out-of-scope action, or whether the impact ratio cleared 0.8. So the catalogue is built from 109 evaluations, each with a metric and a declared threshold, and 90 workflows that produce the evidence a requirement actually needs. 29 common controls connect them to 534 requirements across 51 frameworks, so one run counts everywhere it applies.
The registry covers 51 frameworks, standards and risk lists because that is how many our design partners are actually asked about.
Let the agents do the paperwork. Keep the decisions.
The sentence we test every feature against.
Work with us.
Design partners, contributors, and researchers are all welcome. Start with the repository or a conversation.