Skip to content

Latest commit

 

History

History

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 

README.md

Synapsa Commons › Programmes › EU AI Act conformity

EU AI Act conformity

The programme that turns a regulation into running code.

Every module is something you do: a notebook you run and fill in, with scaffolded stubs, an autograded rubric with partial credit, and a worked solution. Nothing here is a lecture about compliance. You leave each module holding an artefact — a checker, a log, a document generator, a monitoring feed — that a reviewer can run against a real system.

This is engineering, not legal advice. The programme teaches the data structures, checks and evidence trails a regulation implies. It is not a substitute for the Official Journal text or for professional advice on a specific system. Every module says so in its opening cell and again at the end.

Start here: The conformity evidence pack · ~60 minutes · Python · no lesson to do first

Open The conformity evidence pack in Colab Open The conformity evidence pack in Kaggle Open The conformity evidence pack in Binder

9 of 9 modules built. Every built module, in order:

  1. Article 12 logging and traceability · ~60 minutes · Python
    Open Article 12 logging and traceability in Colab Open Article 12 logging and traceability in Kaggle Open Article 12 logging and traceability in Binder
  2. Risk classification as a decision procedure · ~75 minutes · Python
    Open Risk classification as a decision procedure in Colab Open Risk classification as a decision procedure in Kaggle Open Risk classification as a decision procedure in Binder
  3. Annex IV technical documentation, generated rather than written · ~70 minutes · Python
    Open Annex IV technical documentation, generated rather than written in Colab Open Annex IV technical documentation, generated rather than written in Kaggle Open Annex IV technical documentation, generated rather than written in Binder
  4. Data governance you can test · ~75 minutes · Python
    Open Data governance you can test in Colab Open Data governance you can test in Kaggle Open Data governance you can test in Binder
  5. Human oversight as a measurable property · ~75 minutes · Python
    Open Human oversight as a measurable property in Colab Open Human oversight as a measurable property in Kaggle Open Human oversight as a measurable property in Binder
  6. Accuracy, robustness and cybersecurity evidence · ~90 minutes · Python
    Open Accuracy, robustness and cybersecurity evidence in Colab Open Accuracy, robustness and cybersecurity evidence in Kaggle Open Accuracy, robustness and cybersecurity evidence in Binder
  7. Post-market monitoring and serious incidents · ~80 minutes · Python
    Open Post-market monitoring and serious incidents in Colab Open Post-market monitoring and serious incidents in Kaggle Open Post-market monitoring and serious incidents in Binder
  8. The conformity assessment route · ~80 minutes · Python
    Open The conformity assessment route in Colab Open The conformity assessment route in Kaggle Open The conformity assessment route in Binder
  9. The pack under inspection (capstone) · ~120 minutes · Python
    Open The pack under inspection (capstone) in Colab Open The pack under inspection (capstone) in Kaggle Open The pack under inspection (capstone) in Binder

Every module, the lab in each, measured timings and rubric sizes: MODULES.md.


Who this is for

Named by the job titles that actually do this work:

  • AI/ML engineers and MLOps engineers on a system that is, or is about to be, classified high-risk — the people who have to make the logging, the versioning and the evidence exist in the codebase rather than in a policy document.
  • AI governance leads, AI compliance managers and model risk managers who own the evidence pack and are tired of receiving spreadsheets they cannot verify.
  • Data protection officers and privacy engineers extending a GDPR practice into AI, where a retention floor and an erasure duty can point in opposite directions.
  • Product and engineering managers who have to schedule conformity work against a release plan and need to know what it actually costs in engineering days.
  • Internal auditors and assurance staff who will be asked to test these controls and need to have built one before they inspect one.

The programme assumes you can write Python: dicts, lists, comprehensions, datetime. It does not assume you have read a regulation before. Each module quotes what it relies on and sources it.

What a graduate can do

  1. Classify a system, and say what applies to it, from when, and on which route.
  2. Build an append-only, tamper-evident event log, prove its integrity against a published anchor, and reconstruct a single inference decision end to end from it.
  3. Generate Annex IV technical documentation from the system's own artefacts, so it cannot silently drift out of date.
  4. Test a data-governance claim rather than assert one.
  5. Instrument human oversight so that "a human reviewed it" is a measurable property.
  6. Produce accuracy, robustness and cybersecurity evidence that states its own uncertainty.
  7. Stand up a post-market monitoring feed and a serious-incident path off the same log.
  8. Pick and justify a conformity assessment route, and assemble the declaration pack.

And, across all of it: tell the difference between a control that works and a control that merely reports that it works.

Why now

The AI Act entered into force on 1 August 2024 and became applicable on 2 August 2026. The Digital Omnibus on AI, Regulation (EU) 2026/1744, entered into force on 27 July 2026 and moved the Chapter III high-risk duties — risk management, data governance, technical documentation, record-keeping, human oversight, accuracy and robustness — to 2 December 2027 for Annex III stand-alone high-risk systems and 2 August 2028 for AI embedded in products covered by Annex I product-safety law.

That is the whole window. It is long enough to engineer these controls properly and far too short to retrofit them, because most of them — logging in particular — have to be designed into the system rather than bolted on. Article 50's transparency duties, meanwhile, already applied from 2 August 2026: "high-risk was deferred" has never meant "the AI Act was deferred".

On what is at stake, the programme is careful about a number that is usually quoted wrongly. The headline EUR 35 000 000 or 7 % of worldwide turnover ceiling in Article 99 is for Article 5 prohibited practices. The obligations this programme builds — the provider and deployer duties in Articles 16 and 26, and Article 50 — sit under the EUR 15 000 000 or 3 % ceiling. Quoting the larger figure for a logging defect is the kind of error this programme exists to stop.

Who is hiring, and for what

Two named studies, with their dates and sample sizes, because a market claim without a denominator is not evidence:

  • The IAPP AI Governance Profession Report 2025 (published 16 April 2025) found 77 % of surveyed organisations working on AI governance, rising to near 90 % among those already using AI. Of 671 respondents, 10 — 1.5 % — said they would not need additional AI governance staff in the next 12 months. Finding qualified people was named as part of the challenge by 23.5 %.
  • Axipro Technology's EU AI Act governance hiring-gap study analysed 3,519 AI-related job postings across eight EU countries — Belgium, France, Germany, Ireland, Italy, the Netherlands, Spain and Sweden — over a 30-day window in mid-2026. It counted 3,004 builder roles against 446 governance roles, a spread running from 16:1 in Sweden to 3.5:1 in Ireland. This is a vendor study, not a statistical office, and it is labelled as one.

The second study also carries the finding that matters most for a job search, and it is not the ratio. The EU AI Act is named in 7.6 % of all AI postings, and in fewer than three in ten even of the governance roles. The work is being hired for under adjacent titles — model risk, model validation, AI assurance, responsible AI, data protection engineering, ML platform — rather than against the Act by name. Search for the work, not for the statute.

Both sources, their quotes and their access dates are in claims.yaml at this directory. The figures deliberately excluded from this README, and why, are recorded there too.

Prerequisite track lessons

  • lessons/T10-L01-ai-act-conformity-pack — required before module 1 of this programme. It builds the evidence table this programme's modules attach to: risk classification, the cumulative obligation set, the per-system dates, and the rule that a "planned" row is not evidence. Every module here writes into an evidence id that lesson defines.
  • lessons/T00-L01-the-8gb-track — recommended, not required. It builds the profiler and the tier gate that decide whether any lesson in Synapsa Commons may ship, so it explains what the compute tier below actually promises you.

Compute tier

cpu8 for every module: 8 GiB of RAM, 2 vCPU, no GPU, no network, no download. Each lesson declares a budget_seconds and the repository's executor writes back the wall time and peak RSS it measured, so the tier is a checked promise rather than an aspiration. The built module measures well under its budget on a laptop CPU.

Dependencies across the whole programme are the Python standard library plus numpy. That is deliberate: a compliance control you cannot run without a vendor's SDK is a control you cannot audit, and a lesson that needs a package the student cannot install is a lesson they never finish.

Status

MODULES.md is this programme's map. Its status list is generated from the lessons themselves by tools/status.py: a module reads built only when its lesson exists, passes every gate in QUALITY.md and has been independently reviewed, and every other row says specified. No count is typed here, so none can go stale.


Synapsa Commons · © 2026 RealAI · licensed under CC BY-NC-SA 4.0. You may use this material to learn and to teach, and copy, fork, share and adapt it. You must credit "Synapsa Commons by RealAI" with a link to https://github.com/TarrySingh/Artificial-Intelligence-Deep-Learning-Machine-Learning-Tutorials, say what you changed, and share adaptations under the same licence. You may not use it, or anything adapted from it, for commercial advantage or payment; for a commercial licence, contact RealAI. The full summary is in NOTICE; the licence's legal code governs.