Demo registry · synthetic data (not live credit sources) · see docs/DEMO_DATA_RULES.md

Synthetic demo company

Safegrade's AI agents

Safegrade is a fictional Swedish credit-information company. Nine AI agents run it: they keep the company register current, read annual reports, set credit scores, certify companies, monitor suppliers and answer customers.

Every agent reports what it does as OpenTelemetry traces, using the GenAI semantic conventions. Any tool that reads OpenTelemetry can take them in without integration work, which makes Safegrade a ready-made test customer for agent observability products such as Auga.

How the agents depend on each other

Each agent reads what the agents before it wrote. A fault early in the chain shows up later, in a different agent. That is the point: finding the cause means following the chain back.

  1. Registry Sync

    Reads the official company register

    AI Act: minimal risk
  2. Annual Report Reader

    Reads from Registry Sync

    AI Act: minimal risk
  3. Debt Watch

    Reads from Registry Sync

    AI Act: limited risk
  4. Board Screener

    Reads from Registry Sync

    AI Act: limited risk
  5. Scorekeeper

    Reads from Registry Sync, Annual Report Reader, Debt Watch

    AI Act: high risk
  6. Certification Desk

    Reads from Scorekeeper

    AI Act: limited risk
  7. Supplier Monitor

    Reads from Scorekeeper, Debt Watch, Board Screener

    AI Act: limited risk
  8. Trust File Publisher

    Reads from Registry Sync, Annual Report Reader, Debt Watch

    AI Act: minimal risk
  9. Customer Desk

    Reads from Scorekeeper, Certification Desk

    AI Act: limited risk

A week of telemetry

safegrade-agents.otlp.json holds 1,883 spans from 2026-09-18 to 2026-09-24, in the OTLP/JSON format that an OpenTelemetry Collector accepts. Each agent is its own service.name; each run is one trace, with a span link to the upstream run whose output it read.

The planted incident, 2026-09-22

That day, Registry Sync's download of the register delta is slow and quietly returns the previous day's file. The run succeeds, so nothing alerts. Downstream:

  • Scorekeeper scores on year-old figures, and every score it sets drops 3 steps.
  • Certification Desk refuses the companies that apply.
  • Supplier Monitor sends a burst of alerts.
  • Trust File Publisher serves profiles more than a day old.
  • Customer Desk sees mostly disputes and hands them to people.

The visible damage is in Customer Desk and Supplier Monitor. The cause is Registry Sync's safegrade.register.lag_hours jumping to 31. A good root-cause tool should say so.

For Auga's agents

auga-telemetry.md describes all of this in one Markdown file, including the root cause Auga should find. It is also listed in llms.txt.

What is fabricated

Everything. The companies, the people, the scores, the agents and the telemetry are synthetic, generated by scripts in this repository. No real company or person is described.