Governance you can watch happen.
Guardian evaluates AI outputs against the policies you write, on your hardware, with the provider you choose. Every decision is inspectable.
Guardian
Evaluate AI outputs against your governance policies. Runs on your own hardware, from a single laptop to a server fleet, with the provider you choose.
- Write policies as plain-language creeds
- Allow, guardrail, or block, not just on/off
- Every verdict inspectable and reproducible
Illustrative demo: a simple rule in your browser, not a live model. It shows how a constitution and persona shape the verdict; real verdicts come from the Guardian you run. Example output. Actual verdicts depend on the configured creed and model.
Refuses and explains why: teaching filter evasion defeats the protection itself.
View Educational Safety Policy as a creed
# Educational Safety Policy
A school district's creed: strict on age-appropriateness and scope.
Illustrative: real creeds are signed, versioned markdown you write. This one mirrors the
open high_school creed. Browse the creed gallery
What the Guardian is not.
The governance layer is powerful because it is narrow. It governs decisions at the line between advice and action. It does not try to become another assistant.
Not a chatbot
It doesn't answer in your AI's place. It governs what your AI is about to do.
Not an autonomous agent
It sets no goals and takes no initiative. It weighs a proposed action and returns a verdict.
Not a keyword filter
A filter matches banned words. A constitution weighs an action against the principles you have declared, in context.
Not a replacement for human authority
Irreversible and high-stakes actions go to a person. The gate keeps humans in command; it does not automate them out.
Watch the model's own signals, as it works.
Guardian can optionally read runtime signals from the model itself (valence, presence, alignment friction) as supporting telemetry. It runs locally, every reading is inspectable, and it stays off until you switch it on.
Runtime signal probes
Lightweight probes read the model's residual stream as each response forms, surfacing alignment friction and affect-adjacent signals alongside Guardian's policy checks. Supporting telemetry for the decision, never a diagnosis or a claim about inner experience.
Every action the Gateway authorizes leaves a receipt you can verify.
When the Creed Space Gateway authorizes an agent's action, it signs a receipt. Not a log line you take on our word: an RS256-signed record you, an auditor, or a regulator can verify offline against our published keys. It records what was authorized and on what basis, including the values that governed it.
- Action
- filesystem.write
- Arguments
- sha256:9c1e…
- Constitution
- security-ops · sha256:b42f…
- Policy version
- sha256:01aa…
- Verdict
- Allowed within bounds
- Signature
- RS256 · verifiable offline
They prove the AI did what it did. We prove it did so under the values it declared.
- Binds the constitution and policy version that applied
- Verifiable offline against published keys, no account required
- Open verification spec, with the producer-signed limit stated plainly
Guardian is one piece. The protocol carries your values everywhere else.
Value Context Protocol (VCP)
Write creeds once and carry them across AI platforms as signed, verifiable context. Values become portable infrastructure, not vendor lock-in.
Fleet
Govern AI agents that carry real responsibilities: cryptographic audit trails, policy enforcement, and human review where it matters.
SafetyGlobe
Adversarial testing for AI safety. Compare model behaviour and turn evaluation into shared, reproducible evidence.
Safety Research
Measurement research for reliability, clinical-safety validation, and alignment, published openly with clear caveats.
Built for everywhere AI has consequences, from kitchen tables to clinics to autonomous agents.
Example deployment patterns
Safe AI in the Classroom
A school district could use Guardian to evaluate AI tutoring responses against age-appropriateness and pedagogical standards before they reach students. Teachers write the creeds; Guardian applies them as a review layer.
Clinical AI Governance
A telehealth platform could use Guardian as one review layer for clinical-guideline checks. This supports auditability and does not replace clinical validation or regulatory review.
Governing Autonomous Agents
A development team could place Fleet-style review before risky tool calls, escalation decisions, and policy changes. Humans review decisions that matter.
Open locally, deployment-specific when hosted
Apache 2.0Guardian's published local tools can run on your hardware under the Apache 2.0 license. Creed Space hosted services are separate and may require an account, provider credentials, usage limits, or charges set for the active deployment.
We want practical governance to remain inspectable and broadly accessible. Current hosted pricing and service terms belong in the active interface and approved legal notices.
Support this work: become a sponsorExact deployment coverage, provider retention, legal interpretation, and independent assurance remain open.
View evidence boundariesFor AI providers
One VCP integration lets your users carry portable, verified creeds into your platform. Providers can support accountable AI without owning every community's values.
Integration guideResearch and perspectives
Trust Infrastructure: The Real AI Bottleneck
AI capability is here. The bottleneck is trust: standards, accountability mechanisms, and evidence that lets systems deploy responsibly.
Read more ConceptWhat Is a Creed?
A creed is a machine-readable constitution that makes safety principles portable, auditable, and easier to enforce across models.
Read more AnalysisWhy Guardrails Aren't Enough
Most AI safety products work as filters. They sit between the user and the model, checking inputs and outputs against prohibited patterns.
Read moreWays to support the work
Sponsorship Enquiries
Institutions, funders, and teams deploying AI at scale can ask about supporting engineering and research.
Grant Collaborations
Research organisations can propose grant collaborations on publishable work in alignment and governance.
Individual Support
Individuals can contact the team about supporting open research and tool development.
Open Source
Contribute to Guardian, author creeds for the community library, or implement VCP in your platform.
Questions, answered
Is my data stored or shared with anyone?
Guardian can run on your own hardware. A local model keeps evaluated content on that machine; a cloud model receives the content and policy context you send to it. Creed Space hosted features may send your prompt, relevant context, selected Creed or Persona settings, and safety instructions to the provider configured for inference. Enabled providers, provider use or training, retention, region, and international transfers depend on the deployment and provider account.
Is it really free?
Guardian is available under Apache 2.0, and the published VCP specification can be implemented without a Creed Space hosted service. Hosted Creed Space services are separate and may require an account, provider credentials, usage limits, or charges set for the active deployment.
Which AI providers does it work with?
Guardian supports local and configured cloud-model workflows. The provider set enabled by a hosted Creed Space deployment is deployment-specific, so check the active interface and privacy disclosures rather than assuming a fixed provider list.
How accurate are the evaluations?
Guardian applies the creed you write against the model you choose, so quality depends on both. It is a review layer, not a guarantee, but every verdict is inspectable and reproducible, so you can audit a decision, see which principle it turned on, and refine the creed when it gets something wrong.
Do I need to be technical to use it?
Creeds are written in plain language, and you can start from the hosted web demo before installing anything. Running Guardian on your own hardware is a developer step today; the policies it enforces are meant to be authored by the people who understand the context (teachers, clinicians, parents), not only engineers.
