Teacher guide — Grade 12, Unit 8: Capstone with a community client¶

Time: 3 lessons + project (about 5 periods). Track: B/A (project).

#	Lesson	Time
1	Scope and agree	40–45 min
2	Build and audit	40–45 min
3	Hand over and publish	40–45 min
P	Unit project: Community capstone	2 periods

Before you start¶

Materials: Capstone kit in curriculum/grade-12/capstone-kit/ (client brief, data agreement, model card, rubric); notebook notebooks/capstone_template.ipynb.
Read docs/pedagogy/lesson-anatomy.md and docs/pedagogy/ai-use-policy.md.

If you have no computers / no internet¶

A capstone with a simple rule and a printed table is acceptable; call it a 'simple rule' not a model.

Common misconceptions¶

Capstone = biggest model. It is the most useful, honest answer to the client's question.
Student data from real people can be shared freely.

Differentiation¶

Support: give a partially filled table / sentence starters; pair with a data-steward role.
Stretch: ask "what would change your conclusion?" and "who is missing?"

Answer key (check questions)¶

Lesson 1: Scope and agree 1. What does a data agreement cover?
Answer: Purpose, consent, storage, sharing, deletion. 2. Why define success first?
Answer: Otherwise we can't tell if the work helped. 3. May you share a client's data publicly?
Answer: Only if the agreement says so.

Lesson 2: Build and audit 1. What goes in a model card?
Answer: Purpose, data, performance, limits, groups affected. 2. What is the client's 'current way'?
Answer: The practice your model must beat. 3. Why audit before handover?
Answer: To avoid harm to under-served groups.

Lesson 3: Hand over and publish 1. What must be anonymised before publishing?
Answer: Anything identifying people or families. 2. What goes in the README?
Answer: How to rerun, data sources, versions, contacts. 3. Whose consent is required?
Answer: The client, data owners and guardians where students appear.

Assessment¶

Starter quiz each lesson (3 questions, one recalling the previous lesson); unit quiz in quiz.yml; project marked with the rubric in docs/pedagogy/assessment.md; 2-minute oral check per group.

Alignment¶

Enrichment aligned to NEB Grade 12 Computer Science (programming, databases, emerging technologies) and Maths/Statistics.

Answers: practice exercises¶

e.g. "Which Grade 9 students should be offered the Saturday maths support class next term, using last term's anonymised test scores and attendance? Deliverable: a ranked list template and a one-page rule, by Week 6."
Purpose; data fields and period (minimum necessary); consent; storage and access; sharing/publishing; deletion; output ownership/licence (any five).
More visits always find more people; comparing at different capacities confuses "better ranking" with "more work".
Seeds 7/1/2/3: the rule and the model are within about 0.01–0.02 of each other in precision and recall every time; tell the client the simple rule is as good, easier to explain and needs no computer.
Not for deciding who receives care or supplies; not for women outside the district/period of the data; not as an assessment of individual mothers; not to reduce visits for anyone already scheduled; not for insurance or loans.
Do not open/forward it; ask the client to delete it from chats; explain the agreement (only anonymised extracts via the agreed route); inform the teacher; record the incident.
A printed one-page rule and ranked-list template, a simple spreadsheet that works offline, instructions in Nepali, a paper backup — no dependence on internet or the students' laptops.
Recommend using the model only with a safeguard for Mountain-belt women (e.g. the current rule continues for them, or extra review), collect more Mountain data, re-audit after three months; state the gap openly in the model card.
Decline: it exposes and could shame identifiable households, goes beyond the agreed purpose and violates privacy and the class data charter; offer an aggregated, anonymised summary (e.g. payment rate by ward) if the client needs it.
Re-audit every 3–6 months and after any change (new road, new facility, new data system); stop if precision/recall drop below the current rule or group gaps widen; the health post in-charge decides, with a named maintainer.

Lesson 2 (seed 7): current rule 301 visits, precision 0.618, recall 0.741; model at same capacity 0.611 / 0.733.

Real-data exercises: notes¶

R1 Approve the "client" and question; public data still need credits and licence checks.
R2 If the gallery is empty, use published case studies from Nepali data organisations.
R3 A complete synthetic run-through is a valid capstone for schools where no client is available; mark it with the same rubric but note the limits.