Unit 17 activity: Defence, disclosure, and responsible use

Time: about 14 minutes. Checkpoints: cp17-disclosure and cp17-scope-the-claim. You will need: your prompts, outputs and verification results (Units 7, 8 and 10, or Units 9 and 10 on the route without the mathematics), your claim sorting from Unit 16 if your route includes it, and the eleventh row you wrote in the Unit 3 activity: the real piece of your own work you said you would or would not check.

Four short tasks, then the checkpoints. The optional CS extension at the end sits outside the core time budget.

Task 1 — assemble the evidence pack (about 4 minutes)

Open Learner evidence template and fill in items 1 to 6 from work you have already done. You are collecting, not producing. The pack stays with you: it is not submitted and nobody collects it.

That file opens with a table of everything the pack holds and which unit each part came from, generated from course/evidence-pack.yml so that it cannot drift from what the units actually ask you to keep. Work down it rather than from memory.

Two things go in before item 1, and they are easy to forget because you wrote them before the pack existed: the verification sentence from Unit 0, and the task classification table from Unit 1. Unit 18 asks you to look back at the first of them.

Which unit item 3 and item 4 come from depends on your route. On the table-based and code routes they come from Unit 8, the Bernoulli comparison. On the route without the mathematics they come from Unit 9, the citation comparison — the same two items, established the same way, against sources rather than against a table. On the responsible-use route, which is taken beside another path, items 1 to 4 can come from Unit 9, and the rest from that other path.

If you took the Python route, add your self-check record:

python3 selfcheck.py progress --evidence

That prints Markdown. Paste it in.

If you took the table-based route, attach the table you built and the comparison you made instead. This is not the lesser route. The pack is evidencing that you established what correct looked like, compared against it, and can account for the difference — and a hand-built table evidences that as directly as a checker run does.

If item 4 is empty because nothing ever disagreed, say so explicitly and say what you checked. "No discrepancy found, having compared X against Y" is a result. A blank space is not.

Task 2 — write the disclosure (about 2 minutes)

Use the template in the reading. Two or three sentences.

Cover: what you used AI assistance for, at which stage, and what you did about the output afterwards. If you worked from a supplied specimen, say so: no AI system was used, and name the specimen.

Then read it back and apply one test:

Does any sentence in this disclosure argue that the work is correct?

If yes, that sentence is in the wrong statement. Move it to Task 3. A disclosure that has started arguing for correctness has stopped being a disclosure.

Task 3 — write the defence (about 4 minutes)

Two sentences. This is the shortest task and the hardest one.

Sentence one: what your final result is, which specification or convention it is correct under, what check you ran, and what the check showed.

Sentence two: the discrepancy, assumption, or convention you had to resolve, and how you resolved it.

Then run this checklist against what you wrote

• [ ] Does it name a specific check, not "I verified it"?
• [ ] Does it say what the check showed, not just that it was run?
• [ ] Does it name the specification or convention the check was against?
• [ ] Is your evidence independent of the system whose output is under review, rather than another assertion from that system?
• [ ] Does it rely on the model's confident wording, unsupported self-rating, or tool identity as proof? If so, remove that. Tool identity belongs in the disclosure. A confidence measure belongs in a defence only if it has been validated and calibrated for the relevant task, which this activity does not provide.
• [ ] Could someone who does not trust you, and who cannot see your screen, repeat your check from what you attached and get the same result, without asking you anything? (Its steps, inputs and reference must be recorded: that is what makes it a certificate.)

That last question is the real one. A defence is written for a sceptical reader, not a sympathetic one.

Task 4 — turn it on your own work (about 4 minutes)

Everything so far has been about Bernoulli numbers, which is a case chosen so that "correct" is decidable. Your own work is not usually like that, and a method you can only run on the tidy case is not a method you have.

Go back to the eleventh row of your Unit 3 activity: the thing you personally use an AI system for, or are about to. If you wrote "I would not have checked that", this is where that becomes useful rather than embarrassing.

If you have never used an AI system, or have no AI access, take a piece of your own work, or a claim from someone else's work that you rely on, and write the five lines for that.

Write five short lines about one artefact from your own subject or work — one you have generated, one you are about to, or the one you chose above:

1. What it is, in a sentence, and what it would be used for.
2. Which kind of claim it makes. Exact and checkable, source-dependent, a judgement, or a low-risk draft you will rewrite anyway. More than one may apply; privacy and consequence override "low risk".
3. What could fail, and how you would notice. Not "it might be wrong" — the specific failure, and its symptom.
4. The check you would actually run, with the reference it would use, and how long it would take. If no independent reference exists, say that: it is a finding, not a gap in your answer.
5. What you could then honestly claim, and what would remain unchecked.

Add it to your pack as item 9. Nobody collects it. You have now planned a second application of the method, on a case where you had to decide what "checked" would even mean. It becomes a completed verification when you run and record the check.

If you have someone to show it to — a tutor, a colleague, a study partner — a pack is a much better thing to discuss than a finished piece of work, because it shows what you did rather than what you produced.

Task 5 — the checkpoints

python3 selfcheck.py run --unit 17  # or in the browser

Both routes ask the same questions and give the same diagnostics. Do this after Tasks 2 and 3, not before: cp17-disclosure asks what belongs in a defence and cp17-scope-the-claim which sentence the evidence supports, and both are a different exercise once you have written your own.

Optional reflection

Not assessed, not required, and worth five minutes if you have them.

In Note G, Lovelace wrote that the Analytical Engine "has no pretensions whatever to originate any thing" (Lovelace 1843, Note G).

Write one paragraph. Not on whether she was right about language models — that question is genuinely open, and the reading explains why a short confident answer to it should be distrusted. Write instead on this:

Which parts of the work in your evidence pack did you originate, on your own reading of the word, and how would you show it to somebody else?

The interesting answers are usually not about the code.

Optional CS extension

Optional. Outside the core time budget. It does not gate anything, and the course is complete without it.

Take a small routine — twenty or thirty lines is plenty. Your Unit 8 solution works, or anything of your own.

Pick one of three jobs and ask an AI system to do it:

• Translate the routine into another language you can read.
• Review it and report defects.
• Debug it after you have introduced a fault deliberately.

Then produce a test-based account, which is the whole point of the exercise:

1. Before, write down what your tests currently cover. Not what you think the code does — what is actually checked.
2. Run those tests against whatever comes back. For a translation, port the tests too, and note honestly whether porting them was harder than porting the code.
3. Record where the assistance helped, with evidence. A defect it reported that your tests had not caught is evidence. A defect you already knew about is not.
4. Record where it did not, with evidence. Reported defects that are not defects. Real defects it did not report. Changes that altered behaviour your tests did not cover — which is the most instructive category, because it tells you where your test suite is thin.
5. State what you would now add to your tests, and why.

Note the shape of step 5. The most durable value in this exercise is usually not the code that came back. It is finding out which parts of your own program were never really being checked.

Add the account to your evidence pack as an appendix, marked optional.

Before you finish

Your pack should contain all nine items, with the disclosure and defence as separate statements that do not do each other's jobs. Check that once more.

Then, if the work is for assessment, check what your own institution or programme permits and requires. This course does not set those rules.

That is the practical work finished. On the complete route, Unit 18 closes the course by naming the general question behind it, and your pack is what you take into it. On every other route, this is where your route ends; keep the pack.