Turn agent activity into evidence.
Start with the agent you already have. Open-source Guard records covered operations. Bring that evidence into a managed project to review it against approved requirements, evaluate agents, and build toward a governed Gradia Universe with its own tasks, tools, and world behavior.
Starting with human work instead? Workflow Observer assembles approved browser, desktop and service activity into a case. Review the history and business rules, then create a branchable workflow Universe.
1. Install the exact public source release (0.1.0-beta.8) in a Node 20.12+ project.
npm install git+https://github.com/rudycelekli/gradia-guard.git#92a93c610f4d02c3db21fe6a6d526d9bd046fe862. Capture a local process. No account or model provider needed.
npx --no-install gradia-guard run -- node -e 'console.log("Guard is recording")'3. Replace BUNDLE_DIR with the exact directory printed by the run.
npx --no-install gradia-guard verify "BUNDLE_DIR"
npx --no-install gradia-guard inspect "BUNDLE_DIR"Python 3.12+ teams can install the Python wheel from this release. It uses the same evidence format and can be verified independently with either SDK. PyPI publication is pending.
The process wrapper proves lifecycle and covered stdout/stderr byte identities. It captures digests by default. Add gateway, SDK, or runtime instrumentation for stronger coverage. Public npm publication is still pending.
Sign in to bring evidence into a managed project →The audit trail is the beginning, not the product ceiling.
Govern
Record the context, tools, authority, policy, and visible world behind each decision.
Decision and action receipts
Evaluate
Run approved deterministic, model-based, and human-calibrated evaluators over exact evidence.
Evidence-bound verdicts
Benchmark
Freeze representative tasks and compare agents or releases on one governed ruler.
Versioned benchmark editions
Improve
Turn verified failure modes into remediation data and proposed benchmark extensions.
Rights-scoped data products
Certify
Re-run after releases and issue a technical evidence package with drift and regression history.
Release-specific proof
Every claim names its cage.
One line can begin an audit trail. It cannot magically observe an uninstrumented filesystem, hidden memory, or side channel. Gradia records the assurance level in every evidence bundle and refuses claims above it.
| Level | Surface | Can prove | Must refuse |
|---|---|---|---|
| G0 | Process | Covered process dispatch, lifecycle, terminal state, and stdout or stderr byte identities. | No claim about model, tool, file, network, application, or world semantics. |
| G1 | Gateway | Model request, response, provider identity, usage, and dispatch policy. | No claim about tools, files, or world state outside the gateway. |
| G2 | SDK | Model calls, tool definitions, tool actions, application decisions, and declared authority. | No claim that uninstrumented processes, memory, or network paths were absent. |
| G3 | Runtime | Declared files, processes, sidecars, network, clocks, credentials, and side effects. | No counterfactual or restore claim without a complete admitted world snapshot. |
| G4 | Universe | Controlled visibility, evolving worlds, snapshot, fork, restore, replay, and evaluator evidence. | No customer-validity or training-rights claim without independent approval. |
One score. Three independent kinds of judgment.
Gradia does not ask one model to grade another and call the answer truth. Each evaluator owns a declared question, and every disagreement remains inspectable.
Deterministic gates
Required facts, prohibited actions, exact outputs, and hard safety conditions.
Code decides the hard verdict. A model judge cannot turn a failed gate into a pass.
Admitted LLM judges
Quality, judgment, communication, and evidence use where exact code is insufficient.
Every judge is pinned to an approved model, prompt, rubric, threshold, and calibration edition.
Human alignment
Blinded review, disagreement, adjudication, and changes to the rubric or judge.
Human reviewers see the evidence, record why they disagree, and create a traceable alignment signal.
Evidence should be portable.
The Apache-2.0 beta records, verifies, inspects, and compares local evidence without an account. It also simulates a self-digested, deny-by-default model and tool policy before covered dispatches, then maps verified coverage to a local evidence-readiness profile. The account-free capabilities command names twelve local capabilities and the exact evidence ceiling of each. A ready surface is not a compliance or certification claim. The source tag and CI are public; npm trusted-publisher setup and registry release remain separate gates.
Governance becomes an operating system.
Managed G0 ingestion, evidence readiness, AI-system inventory, control workspaces, recurring assessments, custom Gradia Universes, and recipient-scoped auditor packages are production-live. Each package is metadata-only, signed, revocable, and independently verifiable. It proves the evidence relationship it names, not compliance or auditor acceptance. Paying changes availability, never what the evidence proves.
Instrument one agent. Find the workflow worth turning into a Universe.
If your decision depends on a benchmark score, start by reviewing the measurement itself. Explore AI benchmark quality auditing for its inputs, review process and deliverables.