An AI quality system starts with the models in your code.
EN 18286 is the first harmonised standard under the AI Act. Providers of high-risk AI systems who apply it are presumed to meet the Article 17 requirements it covers. Earnie records the models your code ships and the base models they came from, so the evidence behind that quality system starts in the repository.
Request a demoThe dates behind Article 17
Entry into force
The AI Act is published and its obligations start to apply in phases.
In forceAI Omnibus
Regulation (EU) 2026/1744 moves the high-risk dates to December 2027 and August 2028.
In forceEN 18286 cited
The first harmonised AI Act standard, for quality management systems, appears in the Official Journal.
CitedAnnex III systems
The high-risk obligations, Article 17 included, apply to stand-alone systems such as credit scoring and recruitment.
AppliesAnnex I systems
The same obligations apply to AI in regulated products such as medical devices and machinery.
Applies
Sources: Regulation (EU) 2024/1689, the AI Act; Regulation (EU) 2026/1744, the AI Omnibus; Implementing Decision (EU) 2026/2274.
From a model file to quality-system evidence
A quality system has to show what went into an AI system and who signed it off. Earnie builds that record in five connected steps.
Find the models and the AI the code uses
Earnie treats model weights as files to identify, not binaries to skip, and flags the code that calls AI SDKs and services.
Illustrative example
Record what each model is and where it came from
Each model file is matched against the knowledge base. The finding keeps its licence, the base models it was derived from, its datasets and its model card.
- Model file
- models/scorer.safetensors
- Matched by
- Exact file hash, high confidence
- Licence
- Apache-2.0, permissive
- Transform
- Fine-tuned
- Lineage
- mistralai/Mistral-7B-v0.1
Illustrative example
Apply policy before a model merges
AI policies run in developer self-checks and pull request scans, next to the open source and cryptography rules.
A self-check covers only the files it was given. The pull request scan still decides what merges.
Put a name on every decision
A model finding always needs a decision. AI vendors outside your approved list wait for someone to approve them.
- Approvedmodels/scorer.safetensorsM. Ruiz, AI risk owner
- Rejectedmodels/ranker.ptM. Ruiz, AI risk owner
- Wrong identificationmodels/ocr.onnxA. Ito, ML engineer
- Awaiting approvalAI vendor outside the allowlistEnforce approved AI vendors
Illustrative example
Keep the evidence with the release
A snapshot freezes the scan. The AIBOM carries each identified model, and the audit trail keeps every decision as it was made.
purl, SHA-256, licence, lineage, model cardopen source componentsCSV, one row per eventWhere Earnie fits in an Article 17 quality system
EN 18286 is mapped to Article 17(1), which lists thirteen things a quality system has to cover. Pick one to see what Earnie records for it.
Compliance strategy and modifications
Direct evidence- What Article 17 asks
- A strategy for regulatory compliance, including procedures for managing modifications to the high-risk AI system.
- What Earnie records
- Earnie records the model’s lineage, such as fine-tuned, quantised or adapter, and merges declared by the model’s publisher. The Warn on derived AI models template flags a derived model as it arrives.
- What it can’t show
- The compliance strategy itself, and the conformity assessment procedure.
This is our reading of Article 17(1), which the Commission says EN 18286 is mapped to. It isn’t legal advice, and it isn’t based on the clause text of EN 18286. AI model provenance has to be enabled for your organisation.
An evidence layer for your quality system
Identify
Find every model file and AI usage in the code, and match it against the knowledge base.
Govern
Apply model, vendor and licence policies before code merges.
Decide
Approve or reject each model, with a name and a time on it.
Record
Freeze snapshots, export the AIBOM and keep the audit trail.
Whether an AI system is high-risk, and whether a quality system meets Article 17 or EN 18286, is a judgement for the provider and its advisers. Earnie doesn’t classify AI systems, test how a model behaves or report serious incidents. It keeps the model inventory, the policy results and the decision record those processes draw on.
Questions about the AI Act and Earnie
Does Earnie make us compliant with the AI Act or EN 18286?
No tool can do that on its own. EN 18286 describes a quality management system that a provider runs. Earnie keeps part of the evidence that system draws on: which models and AI services your code uses, what they are, which policies they passed and who approved them. Whether a system is high-risk, and whether your quality system conforms, is a judgement for the provider and its advisers.
What is EN 18286?
EN 18286:2026, Artificial intelligence: Quality management system for EU AI Act regulatory purposes, is the first harmonised standard under the AI Act. CEN and CENELEC developed it, and its reference was published in the Official Journal on 14 October 2026 by Implementing Decision (EU) 2026/2274. Providers of high-risk AI systems who apply it are presumed to meet the Article 17 requirements it covers. It is designed to sit on top of ISO/IEC 42001, ISO 9001 or a sectoral quality system such as ISO 13485.
When does Article 17 apply?
After the AI Omnibus, Regulation (EU) 2026/1744, the high-risk obligations, including the Article 17 quality management system, apply from 2 December 2027 to stand-alone systems listed in Annex III, and from 2 August 2028 to AI in products covered by the legislation in Annex I.
Which parts of Article 17 can Earnie support?
In our reading, Earnie gives direct evidence for (a) managing modifications, (b) and (c) design control and quality assurance, (e) technical specifications and (k) record-keeping. It covers part of (f) data management and supports (g) risk management, (h) post-market monitoring and (l) security of supply. Testing and validation, serious-incident reporting, communication with authorities and the accountability framework are organisational processes it doesn’t cover.
How does Earnie identify a model file?
Earnie treats model weights, such as .safetensors, .gguf, .onnx and pickle-based .pt files, as files to identify rather than binaries to skip. It matches each one against the knowledge base and records how it matched and how confident the match is. A file it can’t identify still produces a finding, so an unknown model is visible instead of silently skipped.
What does Earnie record about a fine-tuned or derived model?
Earnie records the model’s lineage, such as fine-tuned, quantised or adapter, and merges declared by the model’s publisher. The Warn on derived AI models template flags these as they arrive, and Require review of models from restricted publishers checks a model’s lineage against publishers you choose.
Can coding agents be held to the same policies?
Yes. A coding agent connected to Earnie over MCP can read the policies in force for a project and check the code it has just written against them before anything is committed. The pull request scan still decides what merges, and the audit trail records the actions an agent takes under its API key’s name.
Does my code leave my environment?
No. Earnie runs on your own server and reads your code there. On SaaS, only fingerprints, one-way hashes of your code, go to the knowledge base we host. On-prem, nothing leaves your network.
Know which models you ship, and who approved them.
See Earnie build a model inventory on code like yours.
Request a demo