SkarpHOW IT WORKS
Anyone can generate a course. Staying right is the hard part.
The upside is real: a rigorous course on last month's regulation, in the language you choose, when nothing comparable exists yet. The catch is that a course made to order is only worth having if it is right. So here is what happens before one reaches you: it is researched before it is written, written against the sources rather than from memory, and re-opened after publication to see whether the field moved without it.
How a course is built
The outline is researched, not recalled
Before a single lesson is written, the subject is researched against current sources — what the field actually covers now, which terminology is in use, what changed recently. A model working from memory alone will confidently teach you the version of a subject that was true when it was trained. Starting from research is what stops that.
Fast-moving subjects are written from the sources
For subjects where being current is the whole point — regulation, standards, security, anything with a version number — the lessons are written against retrieved sources rather than from recall. The material is read at the time of writing, and the sources are kept with the course rather than discarded once the text exists.
Claims that can expire are pulled out and checked
Some statements have a shelf life: a regulation described as in force, a standard called current, a figure given as the latest. Those are identified across the whole course — every lesson, not a sample — and each one is checked against sources rather than against the model's own confidence. A claim that cannot be confirmed is marked unverified rather than quietly passed.
Courses are re-checked after they are published
This is the failure nobody else treats as a failure. A course can be accurate the day it ships and wrong six months later, without a word of it changing, because the world moved. Courses on subjects that move are re-checked on a schedule after publication — the check runs precisely because nothing changed.
Findings go to a person
A check that flags something does not silently rewrite the course. Findings are recorded for review, and a person decides what happens. Automated correction of automated writing is how small errors get laundered into confident ones.
CERTIFICATION PREP
Where the checking goes deepest
Exam material has an advantage the rest of the catalogue does not: an official blueprint to be measured against. Cert courses are built to that blueprint and audited against it, and exam questions are put through independent review — a second model derives the answer without seeing the first one's reasoning, and disagreement sends the question back rather than out. A question that survives is one two independent attempts agreed on.
It matters because the cost of a wrong answer here is not a bad explanation. It is walking into an exam having learned something incorrectly.
DEEP DIVES
When the course is a single document
Some courses are built from one document — a regulation, a research report, a standard — and nothing else. The course is written from the document's own text, walked section by section in its own order. Each lesson is a passage of the source, taught rather than summarised.
Because there is one source, each lesson can be read back against the exact passage it came from — by machine, then by a person — looking for the ways a paraphrase goes wrong: a duty softened into a suggestion, a threshold misremembered, a figure that appears nowhere in the text. What the check flags, a person resolves against the source before the course goes out. A genuine contradiction is fixed before launch, not shipped around.
It matters because the people who reach for a regulation cannot act on a confident paraphrase of it — in its binding provisions, the exact wording is the substance. Here is one provision, three ways.
Ask a general model to summarise a regulation and you get a paragraph — the gist, a few headline duties, the rest compressed away. A deep dive is the other trade: the whole document, walked in order, every provision taught rather than gestured at. The difference shows in a single provision of the Cloud and AI Development Act:
The source — COM(2026) 502, Article 10(1)
“Where data centre capacity is being deployed within the territory of a Member State, that Member State shall designate at least one data centre acceleration zone … within its territory …”
A generic summary
“Member States should set up data-centre acceleration zones to speed up deployment.”
The deep dive
“If a Member State is deploying data-centre capacity, it shall designate at least one acceleration zone — a duty triggered by the deployment itself — and must do so within six months of the Regulation entering into force.”
The summary says less — no trigger, no “at least one”, no deadline — and where it does speak, it drifts: a binding “shall” becomes “should”. The flagging is mechanical; the call is a person's, made against the source. An early draft of this very course made exactly that softening, and the check caught it before it shipped. That is the difference the process is for.
What this does not mean
Every claim on this page is worth exactly as much as its limits are honest, so here they are.
A model writes the first draft
We are not pretending otherwise. The question worth asking of any course is not whether a model was involved, but what was done to catch it when it was wrong. That is what the steps above are.
Depth of checking follows the subject
A course on Roman history and a course on this year's AI regulation do not need the same treatment, and do not get it. Source-grounded writing and scheduled re-checks are aimed at subjects where currency decides whether the material is any use.
Checked is not the same as guaranteed
Verification reduces error. It does not eliminate it. If you find something wrong in a course, tell us — that report is worth more to us than another passed check.
Faithful to the source is not the source
A document course teaches a regulation accurately; it is not legal advice, and it does not replace reading the text where a decision turns on it. What it gives you is an honest, checked account of what the document says — the fastest way in, not a substitute for the thing itself.
Judge it by the material.
Courses are free to listen to. Open one on a subject you already know well — that is the only review that counts.