Agent Incident Record

A neutral registry for what went wrong

Insurance needs loss data. Regulators need comparable reports. Courts need evidence. One standard record can serve all three.

The record

Six questions every incident must answer

  1. What did the system do?

    The action, its time and its context.

  2. Under whose mandate?

    The responsible operator and the limits it set.

  3. Which model and version?

    Base model, vendor and software build (PRV).

  4. What failed?

    The RPS trigger that fired and the observed loss.

  5. Who confirmed it?

    The independent sources and the calculation agent.

  6. What happened next?

    Containment, payout and corrective action.

# Agent Incident Record · illustrative example { "record_id": "AIR-2026-000142", "class": "1 · AI & Software Agents", "operator": "[pseudonymised]", "mandate": { "payment_limit_chf": 5000 }, "provenance": { "base_model": "[vendor/model]", "version": "2026-09-30" }, "trigger": "T-A1 action outside mandate", "observed": { "payment_chf": 18400 }, "evidence": { "log_attestation": "TEE", "timestamp": "RFC 3161", "oracles": "3 of 4 confirmed" }, "status": "confirmed · dispute window closed" }
Registry principles

Trusted because it is neutral

Pseudonymised by default

Operators and individuals are not named in public data. Published statistics are aggregated and anonymised.

Independent

The registry is run by the association, not by an insurer or a vendor, under rules approved by the general assembly.

Useful to all

Quarterly incident reports for regulators and researchers; aggregated access for members to calibrate pricing.

Pilot programme

The registry is in development

We are looking for two or three operators of robots or AI agents to test the record format with us and to help calibrate the first thresholds.