Time: about 14 minutes. Watch the video or read the reading first.
Three short experiments on the thing you cannot see. You need one document that is longer than a page or two — a handbook, a long article, a report, a set of lecture notes — and that you are free to send to an outside service: a public document, or your own notes with nothing personal or confidential in them. The experiment sends the whole of it, and Unit 13 explains why that matters.
If you have no AI access, work on paper with this unit's reading as the document. Part 1: from "What actually goes in", list what would differ between the two askings. Part 2: take this answer, constructed for this activity and produced by no system, to the question "What does the reading say about memory in a new conversation?": Saved memory is off by default and never carries into a new conversation ("The window has an edge"). Open that section, record whether the locator holds and whether the sentence is there, and say which row of the table below it is. Part 3: rewrite the question the way the reading's example does, and say what you could now check. Say in your record that the answer was supplied.
Part 1 — the same question, two contexts (4 minutes)
Ask a question you know the answer to, twice: once at the start of a fresh conversation, and once at the end of a conversation you have already been having about something else.
If your product keeps memory or saved instructions across conversations, note whether they are on: with them on, the fresh conversation is not a clean one.
Record both answers and write one line on how they differ. Length, format, confidence, assumptions carried over from the earlier topic.
Then answer this in a sentence:
If I had only seen the second answer, what would I have concluded about the system?
The point is not that one answer is better. It is that "the same question" described two different inputs, and nothing on screen said so.
Part 2 — the buried clause (6 minutes)
This is the experiment worth doing carefully.
- Take your document and find something stated late in it — the last quarter — that is specific and checkable. A number, a deadline, a condition, a named exception.
- Start a fresh conversation, attach the whole document, and ask a question that can only be answered from that late passage.
- Ask for the locator: the section or page, and the sentence it comes from.
- Go and look.
Record three things: what it said, whether the locator was real, and whether the quoted sentence is at that location and says what the answer said it says.
There are four outcomes and all of them are informative:
| What happened | What you have found |
|---|---|
| Correct answer, correct locator, sentence matches | The answer matches the passage, and you now know how to confirm that cheaply. It does not show the passage reached the model: a public document may also be in training text or found by search |
| Correct answer, no locator offered | You have an answer you cannot check without reading the document. Ask again and insist |
| Confident answer, locator points somewhere else | The characteristic failure of this unit. Note whether anything in the wording warned you |
| The system says the document does not address it | A false negative: you chose a passage that is in the document. The answer did not use it, and this does not say why. It is at least visible, because you can open the passage |
If your document is short enough that everything fits, say so — that is a result too, and it tells you the limit is above the size you work with.
Part 3 — before and after (4 minutes)
Take the vaguest question you asked in Parts 1 or 2 and rewrite it the way the reading's example does: state what would count as correct, ask for locators, and say what to do about absence.
Ask it. Then write two or three sentences on what changed — not about which answer is better written, but about what you can now check that you could not check before.
Keep all three parts. A later unit asks you to record a prompt (Unit 7, or Unit 9 on the route without the mathematics), and after this unit that means recording the whole input: fresh conversation or not, what was attached, what was retrieved.
Now take the checkpoints
cp05-what-reaches-the-system is about what the context actually contains.
cp05-locators is about Part 2's outcome.
python3 selfcheck.py run cp05-what-reaches-the-system
python3 selfcheck.py run cp05-locators # or in the browser
Run them in a terminal from The verification lab and self-check, or use the browser self-check.
Optional extensions
Optional. These sit outside the core study time, are not required for the checkpoints, and nothing later in the course depends on them.
Find the edge deliberately (about 8 minutes). Repeat Part 2 with progressively longer documents until the late-passage question starts failing. You will not get a clean threshold — systems handle overflow differently and the behaviour is not documented — and that is the finding. Record the sizes at which it worked and stopped working, and note that you now know something about your tool that its documentation does not tell you.
Test the format trap (about 4 minutes). Ask for an answer in a very specific format — a table with named columns — about something you can check, and see whether a perfectly formatted answer can be wrong. Compliance with a format is the cheapest kind of compliance to obtain and the easiest to mistake for care.
Marking guidance — open this once you have done the activity, to check your own work
Read this after attempting the activity.
Part 1 — what the difference means
The common finding is that the second answer carries something over: an assumption about your situation, a format from the earlier topic, a level of detail calibrated to a different question. That is the accumulated context doing what it does.
The question at the end of Part 1 is the one that transfers. If you had seen only the second answer, you would probably have concluded something about the system — that it is verbose, or that it assumes too much, or that it is good at this. All of those would have been conclusions about a system drawn from evidence about an input. This is a small instance of a large error, and it is worth having made once in a controlled way.
If the two answers were near-identical, that is a real result and not a failed experiment. A short earlier conversation on an unrelated topic may change little. Try again after a long conversation with strong assumptions in it.
Part 2 — the four outcomes
Nothing here is a pass or a fail. What is being tested is whether you can tell which outcome you got, which requires having gone and looked.
The third row is the one to sit with. When a confident answer cites a location that does not support it, look back at the wording and ask honestly whether anything in it warned you. Usually nothing did — and that is the unit's point rather than a complaint about the tool. The tone of an answer generated from a partial document is the tone of an answer generated from a whole one.
The fourth row is easy to misread. "The document does not address this" sounds like the careful answer, and when a document really is silent it is the right one: the reading's example prompt asks for it explicitly. Here it is wrong, because you chose a passage that is in the document. It is a false negative, and it is the third row's failure with a different surface: the answer did not use that passage, and nothing in it says whether the passage fell outside the window, was not retrieved, or was present and not matched to your question. It is the easier of the two to catch, because the passage you already found disproves it. A statement of absence is a claim like any other, and checking it means looking in the document.
If you used the supplied answer, it is the third row: "The window has an edge" says nothing about memory, and "What follows for checking" says memory can be on by default.
If everything fitted, your document was inside the limit. Say so and move on. Reporting "I could not reproduce the failure at this size" is a finding with a condition attached, which is exactly the form this course wants.
Part 3 — what changed
The answer being looked for is checkability, not quality.
A good response says something like: before, confirming the answer meant reading the document; after, confirming the returned claims means opening three named sections. That is the same claim the reading makes about the handbook example, arrived at on your own material. Whether anything was left out is a separate check: a search of the document.
A weaker response says the second answer was more detailed, or better organised, or more professional. Those are properties of the text. The second answer may also be wrong — the exercise does not establish otherwise — and if it is, you can now find out in two minutes.
The checkpoints
cp05-what-reaches-the-system tests the model of the context window. The
tempting wrong answers all treat the typed prompt as the input, which is the
assumption behind both "prompt engineering as magic words" and "it gave me a
different answer, so the model changed".
cp05-locators tests Part 2's judgement. The trap is treating a plausible
locator as a check. A section number is not evidence until you have opened the
section — it is a claim like any other in the answer, and it is the one claim in
the answer that is cheap to falsify.
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.