Time: about 12 minutes. Read page.md first, at least as far as "The move worth
learning".
You are going to apply the five classes to work you actually have, and then do the one thing this unit is for: take a task that would be expensive to check and move it into a class where it is cheap.
Part 1 — five tasks of your own (5 minutes)
List five things you have used, or would plausibly use, an AI system for. Real ones, from your own study or work. If you have never used one, list five you can imagine reaching for this term.
For each, fill in three columns.
| The task | Class | Where the reference would come from |
|---|---|---|
For the middle column use one of: transformation, grounded answering, generation to a specification, recall, judgement.
For the third column be specific. Not "a source" but which source, and where you would get it. If the honest answer is "I do not know where I would check this", write that. It is the most useful row you will produce.
Two rules that stop this becoming a paper exercise:
- Sort by where the reference lives, not by how the task feels. A task can matter enormously and still be transformation.
- If a task spans two classes, write both. Most real tasks do, and the parts fail independently. "Draft this email and check the deadline I mentioned" is judgement plus recall, and only one of those halves can be wrong in a way that matters.
Part 2 — move one of them (5 minutes)
Look at your five and find the one with the most expensive reference — usually a recall row, and very often the one where you wrote "I do not know".
Write two or three sentences answering this:
What would I have to supply, or write down, to move this task into a cheaper class — and is that thing available to me?
Then say which class it would move to.
Sometimes the answer is that it cannot move. A question about what will happen, or about something nobody has written down, stays where it is. Recording that is a result, not a failure: it tells you the output will be a proposal whatever register it arrives in.
Keep both parts. Unit 17 asks you to assemble an evidence pack, and this table
goes into it alongside the verification sentence you wrote in Unit 0 — the two
things you contribute before the pack's numbered sections begin.
course/learner-evidence-template.md lists everything the pack holds and which
unit each part comes from.
Now take the checkpoints
cp01-where-the-reference-lives gives you five tasks and asks you to match
each to its class — the same judgement you made in Part 1.
cp01-changing-class is about the move you made in Part 2.
python3 selfcheck.py run cp01-where-the-reference-lives
python3 selfcheck.py run cp01-changing-class # or in the browser
Run the Python route from course/lab/, or use the browser self-check. Both give
the same diagnostics.
Optional extensions
Optional. These sit outside the core study time, are not required for the checkpoints, and nothing later in the module depends on them.
Price one check honestly (about 5 minutes). Take the recall row you moved in Part 2 and, before moving it, estimate how long checking it properly would actually take — finding the source, reading the relevant part, confirming the claim as stated. Then do it, and record the real time. The gap between the estimate and the reality is the thing this unit is trying to make visible.
Find a task that resists the move (about 4 minutes). Deliberately look for something you would want to ask that cannot be moved into a cheaper class, and write one sentence on why. Questions about the future, about private facts, and about the edge of what is known are the usual sources. Naming one is worth more than sorting ten easy cases.
Marking guidance — open this once you have done the activity, to check your own work
Read this after attempting the activity.
There is no single correct table, because the tasks are yours. What can be checked is whether each row's third column names something you could actually go and get.
Part 1 — what a good row looks like
A strong row has a specific reference: not "the documentation" but "the
pandas merge documentation for the version I have installed"; not "a paper" but
"the paper the summary cites, at the page it cites".
Three patterns that mean the sorting has not quite worked yet:
| Pattern | What has happened | What to do |
|---|---|---|
| Every row is transformation | You have sorted by what you supply rather than by what would settle it. Attaching a document does not make a question about the wider world grounded. | Ask for each row: could I answer this using only what I handed over? If not, the reference is elsewhere. |
| No row is judgement | Judgement is easy to miss because the answer arrives in the same confident register as everything else. | Look for any row where you asked "which is better", "is this any good", "what should I do". |
| The third column says "I would just check it" | That is an intention, not a reference. Unit 0's sentence had the same failure. | Name the artefact you would open. If nothing comes to mind, that is the finding. |
The rows where you wrote "I do not know where I would check this" are the valuable ones. They are not gaps in your answer; they are the tasks whose checking cost you had not priced, which is the thing this unit exists to make visible.
Part 2 — what moving looks like
The move is almost always one of two things:
- Supply the source. Recall becomes grounded answering. The check goes from "find out what the truth is" to "read this document at this point".
- Write the specification. A vague request becomes generation to a specification. The check goes from "does this look right" to "does this meet the thing I wrote down".
A good answer names the artefact and says whether you have it. "I would need the 2024 version of the module handbook, which is on the intranet" is an answer. "I would give it more context" is not — context is not a reference, and the whole point of the class is what you would compare against.
If you concluded the task cannot move, check that the reason is about the question rather than about effort. "There is no written record of this" is a reason. "It would take a while to find" is a cost, and cost is exactly what this unit asks you to pay deliberately rather than skip silently.
The checkpoints
cp01-where-the-reference-lives uses five tasks distinct from your own and tests
the same sorting. The two confusions it targets are worth naming, because both
survive into later units.
The first is treating anything with a document attached as grounded answering. Attaching your own notes and asking for a summary is transformation: the reference is the notes, and the failure to look for is omission. Attaching a policy and asking what it permits is grounded answering: the reference is the policy, and the failure to look for is a clause stretched past what it says.
The second is treating judgement as generation to a specification. If you can write down what would count as correct, you are specifying. If you cannot, because there is no fact of the matter, you are asking for a proposal — and no amount of precision in the prompt changes that.
cp01-changing-class tests the move itself, and the tempting wrong answer is the
one that adds material without changing where the reference lives. More context
is not a reference.
Your private activity record
Browser storage is not a permanent copy
Progress is kept only in this browser, profile and device. Private browsing, clearing site data, removing the profile, a browser reset, storage eviction or device loss can erase it. Keep important answers and contributions separately.
These notes stay in this browser unless you download a backup or activity log.