API documentation

Send evidence into Guardian through APIs and webhooks.

Guardian connects to your existing AI systems, monitoring tools, and review workflows to turn events, metrics, incidents, and system changes into traceable evidence records.

Book an integration discussion View pricing

Integration overview

How evidence ingestion works

Your stack keeps operating as today. Guardian receives structured signals and turns them into governance-ready evidence you can review and export.

1
Existing AI stack
MLOps, decision logs, monitoring
2
Guardian API / Webhooks
Structured events & metrics
3
Evidence record
Traceable governance object
4
Audit-ready export
Review packages for oversight

Authentication

API keys and Bearer tokens

Programmatic access uses organisation-scoped API keys. Pass your key on every request; ingestion endpoints require a WRITE scope key created in Settings → API Keys.

Authorization header

Authorization: Bearer YOUR_API_KEY

Base URL (production): https://guardian.nordicaiintegrity.dk/api/v1. Metrics ingestion: POST /api/v1/metrics returns 202 Accepted with an ingestion ID to poll until processed.

Event types

Evidence events Guardian normalizes

Map your production signals to a consistent event vocabulary. Guardian turns each into an evidence record your compliance and engineering teams can review together.

decision.created

A production decision or inference outcome is recorded with system context, model version, and review state.

When to sendCredit, fraud, hiring, or triage systems emit per-decision metadata.

metric.drift_detected

Drift, fairness, or monitoring thresholds breach — mapped to alerts and evidence refs without raw payloads.

When to sendPost-market monitoring jobs or batch metric snapshots complete.

review.assigned

Human oversight is assigned or escalated for a decision, model change, or incident follow-up.

When to sendCompliance or operations queues a review in your workflow tool.

incident.opened

A serious incident or production issue is opened and linked to the affected AI system.

When to sendCustomer impact, bias concern, or regulatory notification path starts.

evidence.linked

External artefacts — model cards, policies, monitoring runs — attach to an evidence record by reference.

When to sendTeams reference docs in SharePoint, Git, or MLOps instead of uploading binaries.

export.generated

An audit-ready export package is generated from accumulated evidence and review history.

When to sendInternal audit, regulator request, or procurement due diligence.

Example payload

Webhook / event body

Illustrative JSON for a decision event. Field names align with the evidence record schema below.

{
  "event": "decision.created",
  "system_id": "credit-risk-prod",
  "decision_id": "dec_01842",
  "model_version": "credit-v3.2",
  "timestamp": "2026-05-20T09:45:00Z",
  "review_state": "queued",
  "evidence_refs": [
    "model-card:v3.2",
    "monitoring-run:2026-05-20",
    "policy:credit-risk-review"
  ]
}

Evidence record

Evidence record schema

Normalized records tie systems, decisions, reviews, and external references into one audit trail.

Guardian evidence record fields
Field	Description
`system_id`	Stable identifier for the in-scope AI system (production boundary).
`decision_id`	Unique decision or inference reference from your stack (when applicable).
`model_version`	Deployed model or policy version tied to the event.
`event_type`	Normalized Guardian event (for example decision.created, metric.drift_detected).
`review_state`	queued, in_review, approved, rejected, or escalated — aligned to your workflow.
`evidence_refs`	Pointers to model cards, monitoring runs, policies, or external documents.
`reviewer`	Optional assignee or role responsible for human oversight on this record.
`timestamps`	Event time, ingestion time, and last review update (ISO 8601).
`export_status`	Whether the record is included in a draft or final audit-ready export package.

Data boundaries

Send governance evidence, not raw model data.

Guardian does not require raw model weights, training datasets, or sensitive production payloads by default. Teams send structured metadata, metrics, review states, incident markers, and evidence references — scoped per integration during onboarding.

Security & Privacy →

Implementation

Integration steps

1
Create an integration scope
Define one in-scope production system, event sources, and which metadata fields you can share without raw model data.
2
Send events through API or webhook
Use Bearer-authenticated REST calls (for example POST /api/v1/metrics) or inbound webhooks scoped during onboarding.
3
Normalize events into evidence records
Guardian maps payloads into traceable records — decisions, metrics, reviews, incidents, and linked evidence refs.
4
Review and export evidence packages
Compliance and engineering collaborate in Guardian; export audit-ready packages when oversight requires proof.

FAQ

Common integration questions

Do we need to replace our MLOps stack?

No. Guardian sits alongside your existing monitoring and decision infrastructure. You send structured events and metrics; Guardian adds evidence records and governance exports.

Can we start with webhook events only?

Yes. Many teams begin with a narrow event set — for example metric snapshots or review assignments — then expand to additional event types.

Do you support batch ingestion?

Batch metric snapshots and scheduled jobs are supported via the metrics API. High-volume decision streams are scoped per integration during onboarding.

Can events reference external evidence?

Yes. evidence_refs are designed for pointers to model cards, policies, and monitoring artefacts in your existing tools — not duplicate storage by default.

Do you need raw model data?

No. By default Guardian ingests structured metadata, metrics, review states, and references — not raw weights, training datasets, or sensitive production payloads.

How do engineering and compliance teams collaborate?

Engineering owns ingestion and event mapping; compliance and risk own review workflows, incident registers, and exports. Both work from the same evidence record in Guardian.

Talk to us about evidence ingestion.

We'll help your team map the right events, payloads, and review triggers for your AI systems.

Book an integration discussion