What AIUC-1, ISO/IEC 42001 and the EU AI Act ask for in agent logs, and how NovaFabric's per-run records and bring-your-own control catalogs fit next to them.
Frameworks for AI systems are converging on one practical demand: show what the system actually did. The EU AI Act requires high-risk AI systems to technically allow the automatic recording of events. ISO/IEC 42001 audits look for documented information and monitoring records from an AI management system. AIUC-1, a standard written for AI agents, asks for logs of AI system activity and of tool calls, and lists tamper-evident, independently verifiable logs as supplemental evidence. None of the three prescribes a particular tool for producing that record for an individual agent run.
That is the narrow job NovaFabric does. It is an open-source (Apache-2.0) recorder for AI-agent runs. It does not certify anything, and it does not replace a framework, an auditor or a GRC platform. It produces the per-run records those people ask to see.
NovaFabric is not affiliated with, endorsed by, or certified by the publishers of these frameworks. A mapping shows which recorded evidence exists; it does not certify that a control is met.
EU AI Act (Regulation (EU) 2024/1689). Article 12(1): "High-risk AI systems shall technically allow for the automatic recording of events (logs) over the lifetime of the system." Article 19(1) asks providers, and Article 26(6) asks deployers, to keep those automatically generated logs, to the extent the logs are under their control, for a period appropriate to the system's intended purpose and of at least six months, unless other Union or national law says otherwise. Under Article 113, as amended by Regulation (EU) 2026/1744, these high-risk obligations apply from 2 December 2027 for Annex III systems and from 2 August 2028 for Annex I systems. Logging is also the subject of a draft harmonised standard, CEN-CENELEC's prEN 18229-3, Artificial Intelligence – AI system requirements – Part 3: Logging, listed as under approval in November 2026. Once a harmonised standard is cited in the Official Journal, conforming to it gives a presumption of conformity with the requirements it covers (Article 40(1)).
ISO/IEC 42001:2023 is a management-system standard for organisations that develop, provide or use AI systems. ISO itself does not certify anyone; external certification bodies do. The standard is copyrighted, so we do not reproduce it here; in our own words, an auditor will look for the information the management system keeps and for evidence that operation is monitored.
AIUC-1 is a standard for AI agents published by the Artificial Intelligence Underwriting Company (AIUC), with certification audits run by accredited third-party auditors. Its text is also proprietary. Two of its requirements are directly about records of an agent's behaviour: E015 (in our words: keep logs of what the AI system does) and D003 (in our words: keep the agent from making tool calls it should not, with a record of the calls it did make among the evidence). Under E015, logs that are tamper-evident and independently verifiable are listed as supplemental evidence rather than core evidence: they strengthen an audit, but not every audit requires them. D003 does not ask for it.
The frameworks differ in scope and in who audits them. What they share is the need for an honest record of each run.
nova capture -- <your command> wraps an agent run and writes a directory of
plain files: the model calls made through supported Python SDKs (model-calls.jsonl) and, when tools are
called through MCP or a supported framework adapter, the tool calls (tool-calls.jsonl).nova verify checks the seal against the files later. A change to a
sealed file makes the check fail.nova export-evidence packs a capsule into a signed ZIP that a third party
can check with standard tools (SHA-256 and an Ed25519 signature), without installing NovaFabric or reaching
your infrastructure.With a local signing key, none of this needs a NovaFabric server or a network connection (your agent's own model calls are another matter).
Because ISO/IEC 42001, the AICPA criteria behind SOC 2 and AIUC-1 are copyrighted, you bring the catalog: a small YAML file with the control IDs from your licensed copy and a description of each control in your own words. NovaFabric checks each capsule against it.
id: my-catalog
name: My control catalog
framework: (the framework you hold a licence for)
version: "1"
controls:
- id: <control ID from your licensed copy>
title: Agent activity is logged # your words, not the publisher's
description: Each run keeps a record of its model calls.
evidence_requirements:
- evidence_type: model_calls
description: model-calls.jsonl is present and non-emptynova audit map --profile ./my-catalog.yaml --capsule <capsule-id>The report lists each control as covered, partial, missing or not captured by this version. On NovaFabric's main branch, which becomes the next release, the command also exits 1 when any control is missing, so it can gate a CI job, and a misspelt evidence type, or a control with nothing to check, is refused when the file loads.
The full walkthrough, with a fictional framework you can run end to end, is in the tutorial Map NovaFabric evidence to your own control framework.
model_calls requirement is met by one recorded call; it does not prove
that every call was recorded.nova capture today, so
controls that need them are reported as not captured by this version, rather than as met or as missing.pip install "novafabric @ git+https://github.com/MSKazemi/novafabric@main"The catalog checks above are on the main branch and ship in the next release; after that, pip install novafabric
is enough. Then follow the tutorial linked above. NovaFabric is open source under Apache-2.0:
github.com/MSKazemi/novafabric.
AIUC-1, ISO/IEC 42001, SOC 2 and the EU AI Act are referred to only to describe what they cover. All trademarks belong to their owners.