Pattern: Continuous Assurance Telemetry

Control decisions streamed into one assurance store as they happen, so whether a control works is a live query, not a point-in-time attestation.

By Jorge García Aibar
v0.5.0
Updated 2026-09-25
CC BY 4.0

Layer 05 · Assurance & Continuous Compliance In the chapter 05 catalogue

One working control on the engineering side of AI governance.

Summary: Stream control decisions (policy verdicts, eval results, guardrail actions, identity events) into an assurance store as they happen, so the status of a control is a live query rather than a point-in-time attestation. Trustworthiness becomes a continuously generated signal, not a static certificate¹.

In short

Continuous Assurance Telemetry is an assurance control that streams control decisions, such as policy verdicts, eval results, guardrail actions and identity events, into one assurance store as they happen. It solves the decay of point-in-time evidence: an attestation says a control existed when someone looked, and says nothing about the weeks in between, during which the model was retrained and an agent gained a tool. Use it when the lower layers of the stack already emit structured records and the assurance function is tired of assembling binders before each audit. Each control writes a timestamped record on one common schema: the control id, the subject and version from the registry, the decision, the metric, value and threshold, the obligation reference, an input hash, the actor and a signature. The status of each control becomes a live query over the store. Its illustrative mappings are EU AI Act Art. 72, ISO/IEC 42001, the NIST AI RMF Manage and Govern functions and CSA AICM.

Continuous Assurance TelemetryRuntime signals are checked against controls continuously, and every check writes an evidence record an auditor can read from the assurance store. Continuous means the evidence is produced by the data path, not a quarterly exercise. Generated from the Body of Knowledge.Open interactive diagram (opens in a new tab)

Objectives

Replace periodic attestation with evidence emitted as the system runs, so “is the control working?” is answered by telemetry.

Target users

AI governance engineer, SRE/platform team, auditor.

Impacted stakeholders

Model owners, auditors, regulators, incident responders.

Relevant principles

Instrument the build to produce its own proof; give every control teeth.

Context

A stack whose lower layers already emit structured records, and an assurance function tired of assembling binders before each audit.

Problem

An attestation says a control existed when someone looked; it says nothing about the weeks in between, during which the model was retrained and an agent gained a tool. Point-in-time evidence decays immediately.

Solution

Have each control write a timestamped, structured record to a common assurance store, on one schema. Fix the schema first: a minimum useful evidence record is {control_id, subject (model/agent/system id + version from the registry), decision (pass/fail/allow/deny/alert), metric + value + threshold, failure_mode/obligation ref, input_hash, actor, timestamp, signature}. Normalise every tool’s output into that shape on ingest, so heterogeneous sources compose into one queryable store keyed on the registry id.

Illustrative schema for the evidence record:

{
  "control_id": "guardrail.output.pii.v2",
  "subject": "csa-01@2026-09-18",
  "decision": "alert",
  "metric": "pii_leak_rate", "value": 0.004, "threshold": 0.0,
  "obligation": "EU AI Act Art. 15",
  "input_hash": "sha256:1c7d…",
  "actor": "csa-01",
  "timestamp": "2026-09-18T14:31:52Z",
  "signature": "ed25519:5a…"
}

Two research proposals point at the same idea and are worth watching, not adopting whole: TAIP treats NIST TEVV outputs as reusable AI Assurance Objects that compose across systems¹, and AAGATE operationalises a control plane aligning the NIST AI RMF functions for agents in production²; both are single preprints. Expose the current status of each control as a query over the store.

Consequences

The audit becomes a query and drift is visible in near real time. The cost is building the pipeline and storage, and defining a common evidence schema across tools.

Machine-Readable Evidence (OSCAL); Eval Gate in CI; Runtime Guardrail; Incident Pipeline; Kill Switch / Circuit Breaker.

Maps to: EU AI Act Art. 72 · ISO/IEC 42001 · NIST AI RMF (Manage, Govern) · CSA AICM · Layer 05 Assurance & Continuous Compliance.

Function labels follow the NIST AI RMF³. Mappings are illustrative, not a claim of conformity.

Sources

[1] TAIP: NIST TEVV outputs as reusable AI Assurance Objects; trustworthiness as a continuously generated signal (arXiv 2603.03340; submitted 15 Feb 2026). 2026-02. https://arxiv.org/abs/2603.03340 (verified: primary)
[2] AAGATE: NIST AI RMF-aligned, Kubernetes-native governance control plane for agentic AI (arXiv 2510.25863). 2025-10. https://arxiv.org/abs/2510.25863 (verified: primary)
[3] AI Risk Management Framework (AI RMF 1.0; Govern, Map, Measure, Manage). NIST. 2023-01-26. https://www.nist.gov/itl/ai-risk-management-framework (verified: primary)

Edit this page on GitHub

Cite this pattern

García Aibar, J. (2026). Pattern: Continuous Assurance Telemetry. In AI Governance Engineering: The Thesis & Body of Knowledge (v0.5.0), chapter 05, Patterns. https://doi.org/10.5281/zenodo.22956197. https://aigovernanceengineer.com/patterns/continuous-assurance-telemetry. CC BY 4.0

BibTeX

@misc{aige2026bok,
  author  = {Jorge García Aibar},
  title   = {{AI Governance Engineering: The Thesis \& Body of Knowledge}},
  chapter = {05. Patterns: Continuous Assurance Telemetry},
  year    = {2026},
  version = {0.5.0},
  doi     = {10.5281/zenodo.22956197},
  url     = {https://aigovernanceengineer.com/patterns/continuous-assurance-telemetry},
  note    = {Version 0.5.0}
}

Obligations this pattern evidences

Cases where it applies

Related controls