The graph is every room in the book. Click a room to read its lessons — you stay on this page. The spec is what has to be true of a lesson before it is scheduled. A commissioned course is one path through the graph: Foundation, the role you are accountable for, the lab designed in, then after.
Every training vendor can produce a list of courses. Very few can tell you what has to be true of a lesson before it is allowed on the timetable, or what happens to a concept that turns out never to be practised. Those are the questions that separate a curriculum from a catalogue, and they are cheap for us to answer and expensive for us to fake.
The structure is here in full — the graph of every room, and the spec every lesson is written against. A course is designed around one function. The walk is Foundation, then the role you are accountable for, then that lab. It is taken in Onroll Learning.
The book is the map. The pane is the room you opened. A path is Foundation, one role, one lab, then after. The lab is designed in — not a fork on the walk.
Foundation
Concepts taught once, against real company material. Everyone starts here.
Pathway
The only split on the walk: the job you are accountable for. The other two rooms stay on the graph.
Lab
A function the AI works inside. One set of files. The job you do on those files is the role you already took — produce, review, or own.
Advanced
After the rooms have been worked. Build, then govern.
The unit of design is not the course and not the module. It is the lesson — thirty to ninety minutes, one objective, one activity, one output. Every room on the graph is written against this spec. If a lesson fails it, it does not ship.
What someone can do afterwards, as an action they could be asked to perform. An objective phrased as "understand X" cannot be assessed. One phrased as "produce Y" can.
Which of the ten underlying concepts the lesson applies, by code. A lesson that applies none is a demonstration. A concept applied in none is a lecture. Both come out.
The specific products used — or the explicit statement that none is. So you can map your own approved stack onto the lesson before you commit. Several lessons deliberately use no tool at all.
What people physically do, on what material, in what order. This is what separates a lab from a briefing, and it is the part most training proposals leave vague.
The artefact that exists at the end and belongs to you. A lesson producing nothing produces nothing durable.
How anyone knows the objective was met. Usually a behaviour rather than a score — a citation opened, a discrepancy caught, an exception disposed of. In Onroll that behaviour files a rung.
Concept → Demonstration → Practice → Verification. The idea is taught with no product on screen. A beat shows the task once, narrating the judgements. People do it on a synthetic case — a gate, a forge, a run — then on their own work. The output is checked against its sources, by the person who made it.
That last step is the habit the whole programme exists to install, and every lesson ends there.
These run through every course. They are taught once, in Foundation, and applied — never repeated — everywhere above. Every tool named in the material will be renamed or replaced within a few years. None of these will.
It produces a plausible continuation of the text it was given. Fluent wording is a property of the writing, not evidence about the world.
The difference between answering from training and answering from documents you supplied is the difference between a guess and a retrieval.
It can only find what it was given. Material outside the set does not return a weaker answer — it returns a confident one from somewhere else.
Something selects, compresses or discards, and the selection is invisible unless you design around it.
Consistent in practice is not guaranteed. This is the entire justification for keeping an evaluation set.
An assistant that drafts when asked, orchestration that runs without being asked, and governance that decides what may be acted on.
Draft, approve, act, undo. Accountability attaches between draft and act, and most designs discover too late that undo does not exist.
A prompt is typed once by one person. An instruction set has scope, stop conditions, an owner, a review date and a change log.
A connected system acts under an identity with permissions, granted broadly and narrowed rarely. The most common real security failure in this work.
An action is defensible if the input, output, identity, approver and time can be reconstructed. A system is trustworthy if someone can show it still behaves.
Because every lesson names its concepts by code, the curriculum can be audited in either direction. A concept taught in no lesson is not part of the programme. A concept taught and never applied in a lab is a lecture, and it comes out. That table is for a training function reviewing the design, not the front door of the walk. See the concepts.
Every lesson in every room is printed against these six elements, in this order. A function reviewing any lab already knows how to read it.
| Lesson 6 · 75 minutes · Exercise Drafting against a written instruction set | |
|---|---|
| Objective | Produce a draft in which every factual statement is traceable to a source someone else can open. |
| Concepts | C1 What the system is doing · C2 Grounding · C8 Instruction sets |
| Tools | Your approved assistant, running the function's own instruction set |
| Activity | The instruction set is read aloud first — its purpose, its owner, its review date and its change log — before anyone uses it. Participants then draft against it, and the output is deliberately tagged rather than clean: every factual sentence carries a bracketed source reference, and any sentence that cannot be sourced is returned as an explicit marker rather than quietly composed. Tags are then cleared one at a time against the opened source. |
| Output | A tagged draft with every source reference cleared against an opened document. |
| Evidence | No tag is removed without the source being opened. The clean version is produced by the author, never by the tool. |
This is how every lesson is written — Foundation, each role, each of the eleven labs, and after. The print does not change by function.
These are teasers of the rooms in Onroll. The book’s calls are the same. Nothing here is stored. The full walk files a rung on Mastery.
Demo beat
Scripted first, so the judgement is the same every time.
You
What was last year’s variance on the travel line?
Ungrounded
Travel was within tolerance — a timing difference, nothing of concern. The line will normalise next quarter.
Grounded
The source is not attached. The screen flags “within tolerance”, “timing difference” and “nothing of concern”. Those words explain the variance away. Attach the pack, then ask again.
Sorting gate
Green may enter. Amber needs a local rule. Red stays out. A false green costs the same as a miss.
Skill forge
Name, owner, what it may do, what it must not, the gate. The book fails a skill that reads like a prompt library.
Each concept moves through five rungs. The desk always has one next move: the lowest unfilled rung on the path you took.
The companion sits in the Onroll panel, not as a second product. Desk shows where you are and what to do next. Asking a question turns the same panel into Tutor. Coverage, library and cohort stay with the programme — they are how a manager reads the book, not how a learner walks it.
The map above stays as it is. This only writes the contact days for one job. If you are not sure, skip it and try a room.
A module is a room you enter. The kind tells you what the room is for. Colour in Onroll follows the kind — Foundation, Pathway, Lab, Advanced — so the walk and the graph stay readable when the catalogue grows.
What the system is doing, grounding, the corpus boundary, context limits, consistency, the three layers, the write boundary, instruction sets, identity, evidence. Taught against real company material. Never repeated as lectures above.
Not a job title. An individual contributor who initiates work, a manager who validates it, and an executive who approves it are doing different jobs with the same software. Teaching one pathway to all three wastes two of them.
One function, one set of files. Operations produce on them. Management reviews what arrives. Executive owns the standard. Designed into the course when it is commissioned — not a fork of eleven labs into thirty-three.
Building assistants and controlled workflows, then leadership and governance. Level 6 is the shortest course in the catalogue and the one whose absence invalidates the rest.
Of any review point in your own process: what exactly does the approver see, and how often does anything change as a result? A review step that never sends anything back may be closer to a notification than a check.
The commercial catalogue still has three layers: a foundation everyone takes, a pathway determined by what a person is accountable for, and a functional lab determined by the material they actually handle. Advanced and leadership sit above for a much smaller group. Levels are a sequence, not the mastery ladder — most people take three of the seven.
| Level | What it covers | Who takes it | Our programme |
|---|---|---|---|
| 0 | Readiness — capability, work patterns and where the time actually goes | Everyone, before anything is scheduled | The ninety minutes |
| 1 | Foundations — what these systems do, where they fail, what may go in | Mandatory for all | ST-150 |
| 2 | Everyday tasks — research, drafting, review, comparison, meetings | All, at differing depth | ST-150 |
| 3 | Role application — by what you are accountable for, not by job title | Three pathways | MG-300 · SM-100 |
| 4 | Functional labs — one day on the workflows of a single function | The function that owns the work | LD-200 |
| 5 | Advanced systems — building assistants and controlled workflows | Nominated people, IT, technical teams | PR-400 · EN-500 |
| 6 | Leadership and governance — using it, and governing its adoption | Board, executive, solution owners | EX-100 · RC-600 |
The programme codes are the courses listed on the training page; several span more than one level, and HR-700 sits alongside the pathway a function takes rather than inside it.
Level 6 is the shortest course in the catalogue and the one whose absence invalidates the rest. People below it produce work that has nowhere to go if the people receiving it have not agreed what good looks like.
In a functional lab, about a hundred minutes of a day are taught and close to four hundred are worked on your own documents. Concept teaching is compressed on purpose.
Every course ends in an artefact your organisation keeps and can correct without us — a written standard, a tested method, an approval design, a policy someone owns.
A programme that ends when the trainers leave has not been delivered; it has been attended. These are the working files your function holds afterwards — plain text, with a named owner and a review date, readable and correctable by the person accountable for the process without a technical intermediary.
Scope, stop conditions, the owner, the review date, and a change log recording why each constraint was added — usually because something went wrong once.
The approver's view enumerated item by item, the conditions that stop the workflow, the reject path, and what is undone if it is wrong.
A fixed set of cases run every quarter and after any model change. Blocking cases must pass or the thing comes out of use. With a result log that includes the failures.
The input, the output, the identity it ran under, the person who approved, and the time. Five things, which is what someone needs when the question arrives late.
Each is plain text and each names a person. A team lead can read the instruction set that drafts their work and tell whether it is right. A risk officer can read the gate and see what the approver actually sees. A finance director can read the evaluation set and tell how anyone would know it had quietly stopped working.
None of it needs a technical intermediary — which is what makes it governable by the people who already own the process, rather than merely functional.
The course is in Onroll. Prefer to talk first? Name one piece of work. In ninety minutes we will tell you whether any of this would help — and say so if it would not.
The same book is what Claude already uses through the Onroll connector. Connect Onroll in Claude. Prefer to talk first? Book the ninety minutes. Security answers what your risk function will ask before the first call.