Skip to content

Latest commit

 

History

History
176 lines (137 loc) · 7.99 KB

File metadata and controls

176 lines (137 loc) · 7.99 KB

ACR Telemetry Schema Specification

Version: 1.0
Status: Stable
Last Updated: March 2026

This document defines the canonical schema for telemetry events emitted by ACR-aligned control planes. It supports Pillar 4: Execution Observability, audit trails, drift detection input, and compliance evidence generation.

Purpose and Scope

  • Purpose: Provide a consistent, machine-readable structure for all runtime events related to AI system actions, policy decisions, drift alerts, containment actions, and human interventions.
  • Scope: Events emitted by the ACR control plane (or equivalent governance layer) during request processing. Out of scope: model training telemetry, pre-deployment validation logs, and non-ACR application logs.
  • Consumers: Observability pipelines, audit stores, drift detection services, compliance reporting, and incident response tooling.

Event Types

Every event MUST include an event_type from the following set:

event_type Description
ai_inference Standard model invocation (request → policy checks → model/tools → response)
policy_decision Policy evaluation without inference (e.g. pre-flight check, re-validation)
drift_alert Anomaly or drift detection event (behavioral baseline deviation)
containment_action Automated or manual containment response (throttle, restrict, isolate, kill)
human_intervention Manual override, approval, or break-glass action

Schema Overview

Events are JSON objects. Field names use snake_case. Timestamps use ISO 8601 with timezone (e.g. 2026-03-16T14:22:01Z). All UUIDs are RFC 4122 version 4 where applicable.

Top-Level Fields

Field Type Required Description
acr_version string Yes ACR schema version (e.g. "1.0"). Semantic versioning for backward compatibility.
event_id string Yes Unique event identifier (UUID v4 recommended).
event_type string Yes One of: ai_inference, policy_decision, drift_alert, containment_action, human_intervention.
timestamp string Yes ISO 8601 timestamp with timezone.
correlation_id string No Links this event to a request or trace; should be consistent across spans of the same action.
agent object Yes* Identity and purpose binding. *Optional only for non-agent events (e.g. system health).
request object No Request context; required for ai_inference and often for policy_decision.
execution object No Execution details (duration, tool calls); typical for ai_inference.
policies array No Policy evaluation results; present when policy checks were performed.
output object No Output metadata (tokens, cost); for inference events.
metadata object No Environment, drift score, and other extensible attributes.

Object: agent

Field Type Required Description
agent_id string Yes Stable identifier for the AI system or agent instance.
purpose string Yes Operational purpose or scope (e.g. customer_support, data_analysis).
model object No Model identifier and optional version/vendor.
risk_tier string No Risk classification if defined (e.g. low, medium, high).

Object: request

Field Type Required Description
request_id string No Client or gateway request ID for correlation.
input object No Sanitized or summarized input (avoid logging full PII); size limits apply.

Object: execution

Field Type Required Description
duration_ms number No End-to-end duration in milliseconds.
tool_calls array No List of tool invocations (name, params summary, result summary).
error string No Error message or code if the execution failed.

Object: policies[]

Each element represents one policy evaluation:

Field Type Required Description
policy_id string Yes Policy or rule identifier.
decision string Yes allow or deny.
rule_id string No Specific rule that fired.
transformations number No Count of transformations applied (e.g. redactions).

Object: output

Field Type Required Description
tokens object No Input/output token counts if available.
cost object No Cost or usage metadata if available.
redacted boolean No Whether output was redacted or truncated for logging.

Object: metadata

Field Type Required Description
environment string No Deployment or environment label (e.g. production, staging).
drift_score number No Current or latest drift score in [0.0, 1.0] when applicable.
containment_tier string No For containment events: e.g. throttle, restrict, isolate, kill.
approver_id string No For human_intervention: identifier of the approver (opaque or hashed).

Additional custom fields MAY be added under metadata or as top-level extensions; implementers SHOULD namespace custom keys (e.g. vendor_*) to avoid clashes with future ACR schema versions.

Validation Rules

  1. Schema version: acr_version MUST be present and parseable. Consumers MAY reject unknown major versions.
  2. Event size: Events SHOULD stay under 10 KB to avoid pipeline and storage bloat. Redact or truncate large payloads (e.g. request.input, tool call payloads) and document redaction policy.
  3. Timestamps: timestamp MUST be ISO 8601 with timezone. Prefer UTC.
  4. Async emission: Telemetry MUST NOT block the action pipeline; emit asynchronously where possible.
  5. Integrity: For audit logs, use append-only storage and optional hash-chaining as described in Pillar 4.

Versioning

  • Major (e.g. 1.0 → 2.0): Breaking changes (removed or renamed required fields, changed types).
  • Minor (e.g. 1.0 → 1.1): New optional fields or event types; backward compatible.
  • Patch: Clarifications and non-structural edits only.

Implementations SHOULD include acr_version in every event so consumers can route or transform by version.

Examples

Minimal ai_inference event

{
  "acr_version": "1.0",
  "event_id": "550e8400-e29b-41d4-a716-446655440000",
  "event_type": "ai_inference",
  "timestamp": "2026-03-16T14:22:01Z",
  "agent": {
    "agent_id": "customer-support-01",
    "purpose": "customer_support"
  },
  "request": {
    "request_id": "req-abc-123"
  },
  "execution": {
    "duration_ms": 385
  },
  "policies": [
    { "policy_id": "pii_redaction", "decision": "allow" }
  ],
  "metadata": {
    "environment": "production",
    "drift_score": 0.12
  }
}

containment_action event

{
  "acr_version": "1.0",
  "event_id": "660e8400-e29b-41d4-a716-446655440001",
  "event_type": "containment_action",
  "timestamp": "2026-03-16T14:25:00Z",
  "correlation_id": "trace-xyz-789",
  "agent": {
    "agent_id": "customer-support-01",
    "purpose": "customer_support"
  },
  "metadata": {
    "environment": "production",
    "drift_score": 0.72,
    "containment_tier": "restrict"
  }
}

References


ACR Framework v1.0 | Home | Pillars | Implementation Guide