Fall 2026 · Aug 3 – Dec 15 · 85 instructional days
Computer Science I
Course Calendar
An AI-building course · governed by the Identity Charter (CS1_TO_APCSA_ONRAMP)
No school: Sep 7 (Labor Day) · Sep 8 (PD) · Oct 8 (PD/Wellness) · Oct 9, 12–13 (Fall break) · Nov 23–27 (Thanksgiving) · Dec 11 (PD/Wellness)
Staggered start: CS I is 10th and 11th grade only; those grades begin Aug 4, so Week One runs Aug 4–7. Aug 3 is a student-free work day for this course.
Testing windows: Q1 benchmark Sep 28 – Oct 2 · Fall LEAP Dec 1–18 · Q2 benchmark Dec 3–10
Spring, for reference: ACT Mar 9–12 · AP exams by May 19 · last student day May 19.
CS I is the course about the power of AI to understand and wield technologies beyond what you could build by hand, and to create genuinely useful things. Students direct AI, read what it made, judge it, iterate — builder, not consumer. No student is expected to write a line of code from scratch. The AP CSA on-ramp rides underneath in silence: the reading/judging/specifying work happens to seed spring's thinking, but that is a byproduct, never the point. Below, the AI-building work is the course; the (byproduct) notes are margin only — if a week's real purpose ever becomes the margin note, it fails the Charter's §4.1 silence gate. Resource: Code.org AI Foundations; AI Essentials; JM Ch 1–8 for concept support only.
Module 1 — Reading the Data
Aug 3 – Sep 4 · 25 daysOnboarding and baseline. Students establish course routines and produce the baseline dataset: cognitive (knowledge), attitudinal, and behavioral. Each instrument pairs fixed-response items with open-ended items. Students can state what the survey is for, that they will see it again in December, and that their own responses become the dataset the class analyzes in Week 4.
Byproduct: the Module 1 dataset.
day-by-day outline
From survey to class AI policy · Aug 4–7, 2026 · Module 1: Reading the Data
Take the survey. Find what it wouldn't let you say. Fix it. Run the better one. Write the policy from what it shows. Nine class days across Weeks One and Two, one destination — the policy cannot be written without the data work. Every technical decision has a consequence a student can see: response types matter because you have to aggregate 75 people; the dictionary matters because you're about to combine answers and need to know what the values mean; the mean matters because a 3.4 is about to justify a rule.
Two instruments
Named for students on Day 1, with the distinction said out loud: what we measure to learn about ourselves, versus what we measure to measure ourselves.
| Instrument | Purpose | When |
|---|---|---|
| Attitudinal & behavioral survey (v1) | This is our dataset — and the evidence base for the policy. | Administered Day 1, returned Day 3 |
| Cognitive baseline | This is how I find out where to start. Graduated MS→HS, five buckets; SLT anchor and December matched pair. | Closing 30 min, Days 1–3 |
Four 90-minute blocks. Two segments per block, content shorter than lab, each running I-do / we-do / you-do. Baseline occupies the closing 30 minutes of Days 1–3; Day 4 runs the full 90.
Day 1 · Aug 4 — What we are and what we're making
| Time | Segment |
|---|---|
| 0–15 | Who I am and why I'm here. Short — the course earns credibility by what it does. |
| 15–25 | How this class runs. Two segments per block. Lab notebook — I-do, model setting one up. Framing the work, not formatting a header. |
| 25–50 | Survey v1. Who-are-you items collected as categorical variables, not icebreaker. Framed explicitly: you are building the dataset this unit works with. |
| 50–60 | What did we just make? Rows and columns. Respondent, variable, response. First vocabulary, drawn from the thing in front of them. |
| 60–90 | Cognitive baseline, part 1. |
Standard: MS-DAT-DC-22 (data vs. metadata, lightly)
Note. The 50–60 segment is doing more than it looks — ten minutes to convert a thing they just filled out into a structured object, and it is the entire setup for Day 2. If anything overruns, this is what gets squeezed, and it is the one that can't afford it.
Day 2 · Aug 5 — Describing the data you collected
| Time | Segment |
|---|---|
| 0–10 | Content: what a data dictionary is. I-do two entries against our own survey. Name, type, allowable values. One wrong on purpose, fixed aloud. |
| 10–25 | We-do — two items chosen because they split the room. Structured disagreement: two proposals up, class argues, class decides, everyone writes the decision in their own notebook. |
| 25–55 | You-do — each student writes the dictionary for a stated set of remaining items. |
| 55–60 | Notebook close. One line: what was hardest to specify? |
| 60–90 | Cognitive baseline, part 2. |
Standard: HS-DAT-DC-22 (CORE — primary target of the day)
The we-do items are the two without an obvious answer — chosen to split the room. Candidates:
- Prefer not to say — allowable value, or null? Two defensible answers with different consequences on Day 3.
- A Likert item — ordinal or quantitative? Can you take a mean of it? Plants Day 4 two days early.
- Open response — what are allowable values for free text? The honest answer is uncomfortable.
- Grade level — ordinal or nominal, and does it matter?
Why their own survey. A dictionary written against someone else's dataset is copying. Written against an instrument they personally answered, the ambiguities are ones they felt.
Binding constraint. Item count sets the lab. Four items go in the I-do and we-do; the rest land in 30 minutes on a skill learned twenty minutes ago. Define "done" as a stated subset, with the rest as extension — Day 3 needs the dataset, not the whole dictionary.
Day 3 · Aug 6 — Your data comes back
| Time | Segment |
|---|---|
| 0–10 | The raw sheet, on screen. Their responses, unprocessed. Deliberately unglamorous. |
| 10–25 | Content: categorical vs. quantitative. I-do one of each from their data, we-do one. Why the distinction determines what you can do next. |
| 25–55 | You-do — sort, filter, group, summarize. Each student finds one thing that surprises them and writes down why it's surprising. |
| 55–60 | Two or three surprises shared aloud. |
| 60–90 | Cognitive baseline, part 3. |
Standards: S1-DSC-AM-04, S1-DSC-MI-07 (REACH); MS-DAT-DC-23 (SCAFFOLD)
Depends on the demographic decision from Day 1 — grouping requires grouping variables, or this lab is sorting a list.
Day 4 · Aug 7 — Making it show something, and making it lie
Full 90. No baseline.
| Time | Segment |
|---|---|
| 0–10 | Content: what a chart is for. I-do — one variable charted well, then charted badly. Same data. |
| 10–20 | We-do — critique a misleading chart together. Truncated axis, or a mean over a bimodal split. What the mean conceals. |
| 20–50 | You-do — two charts from the class data, one honest and one deliberately misleading. Same variable. |
| 50–65 | Gallery. Read each other's. Identify the manipulation without being told. |
| 65–85 | Course identity. Grounded in what they just did — you made data mislead on purpose, which means you now know it's made. Builder, not consumer. |
| 85–90 | Where the semester goes. |
Standards: HS-DAT-DI-25, HS-DAT-DI-26 (CORE); S1-DSC-VZ-10, S1-DSC-VZ-11 (REACH)
The identity segment lands here rather than Day 1 because it arrives as a conclusion they helped reach instead of an assertion. It is also the week's only AI-adjacent moment in a module whose CORE standards are all data — worth being deliberate that the thesis gets planted.
v0.1 — Week One designed to the block. Timings are targets, not a script.
The machine and where it lives. Students explain how a computer computes, and distinguish VRAM, dedicated GPU, iGPU, and unified memory well enough to say what a given machine can and cannot run.
Students trace one request twice — the same prompt to the classroom's local AI and to a commercial cloud AI — naming each hop along both paths, and account for why the AI in this room does not run on the Chromebook in front of them.
Students compare what each endpoint is permitted to do with what arrived: what the local stack stores (access tokens, no PII, nothing persisting between sessions) against what a commercial provider's terms permit. Students distinguish nothing persists from requests are unlinkable to a person and say which claim each system actually makes.
day-by-day outline
Aug 10–14, 2026 · Module 1: Reading the Data
Design not written — intent and shape below. Five blocks, same two-segment structure. No baseline; those thirty minutes return to instruction. Block timings are absent deliberately, not omitted.
Day 5 · Aug 10 — What couldn't you say
The critique prompt is "what couldn't you say," not "what was wrong with it." The first routes to foreclosed answers — the actual mechanism. The second routes to "too long" and "boring."
A policy survey written by the person who holds the policy asks about the options that person already considered. How often do you use AI on homework has already decided that the interesting variable is frequency. A student whose honest answer is "I use it to explain things I didn't understand, never to produce work" has nowhere to put that. They pick a number, the number gets averaged, and the policy regulates frequency. Nobody was careless — someone modeled the respondent on themselves.
Day 6 · Aug 11 — Why the questions have to be shaped
The constraint arrives from the goal, not from the teacher:
Scale → aggregation → comparability → constrained response types.
To say anything about 75 students you must combine answers. Combining requires comparable answers. Comparable answers require constrained types. Free text doesn't aggregate — which is why the item that feels most respectful of a student's view is the one that can't drive a policy.
| Type | Order | What you can do |
|---|---|---|
| Nominal | none | Count |
| Ordinal | ordered, unequal gaps | Median, distribution |
| Interval / count | equal gaps | Arithmetic |
| Free text | — | Read it, or use a tool |
The trap to plant: you can compute a mean over Likert. It just doesn't mean what people think. A 3.4 on a five-point scale can be a room split down the middle with nobody at 3.4.
The free-text item stays. It captures what the constrained items foreclose, and 75 open responses are the reason Module 1 needs a computational tool at the end of August — a motivation now student-generated rather than assigned.
Day 7 · Aug 12 — Build v2
Students author replacement and additional items. Revision against a concrete object, not a blank page.
The second lesson, and the guard against cynicism: v2 has its own bias, authored by them. Good instruments are not unbiased; they are honest about what they assume. If v2 goes out unexamined as "the fair one," they learn that instrumentation is a problem you solve once.
Day 8 · Aug 13 — Run v2
Administration is short. The rest of the block is analysis — a dictionary for the new items, then the same sort / group / summarize work against a dataset they designed.
Day 9 · Aug 14 — Write the policy
The class drafts the AI policy from v2's results — v0.1, with a stated review date.
Two instruments measuring the same construct, differing by known design changes, both authored in the room. Which items moved, and did they move because opinion changed or because the question changed?
Flag: this day has no standard attached. It is real student work with nothing on the crosswalk — either map it or name it as deliberate non-standards time, per the roadmap's what we are not covering convention.
Standards landing (Weeks 1–2)
| ID | Tier | Day |
|---|---|---|
| MS-DAT-DC-22 | SCAFFOLD | 1 |
| HS-DAT-DC-22 | CORE | 2 (revisited 8) |
| S1-DSC-AM-04 | REACH | 3 (revisited 8) |
| S1-DSC-MI-07 | REACH | 3 |
| MS-DAT-DC-23 | SCAFFOLD | 3 |
| HS-DAT-DI-25 | CORE | 4 |
| HS-DAT-DI-26 | CORE | 4 |
| S1-DSC-VZ-10 | REACH | 4 |
| S1-DSC-VZ-11 | REACH | 4 |
Eight of eleven Module 1 standards, including all three CORE that were pending confirmation — if Week One runs, the Module 1 presumption resolves. The remaining three (HS-DAT-DC-23, HS-DAT-DC-24, S1-DSC-CC-02) belong to MuggsOfSources, Aug 31–Sep 4.
v0.1 — Week Two shaped, not scripted. Block timings deliberately absent.
How a model gets made. Students describe pretraining, post-training, freezing, and inference as distinct phases, and explain parameters and tokenization at the level of what the model stores and what it reads.
Students trace a prompt from text to tokens to output.
Students explain why training is a one-time capital cost and inference is a permanent per-query cost, and connect that asymmetry to why a handful of organizations build models while many deploy them.
Prompt-writing begins here — low-dose and recurring, not a unit.
Data, and their own data. Students classify data types and write a data dictionary for the Week 1 survey schema, including a limitations section naming who is not in the sample, what the instrument could not ask, and what nonresponse does to the result.
Students perform thematic coding by hand on the class's open-ended responses: familiarization, line-by-line coding, collating codes into themes, checking themes against extracts. Students locate the response that fits no theme and decide — force it or name it — then write the criterion they used.
Students then submit the same responses to a model with no special instruction, compare against their own codebook, and write a prompt that protects the outlier using the criterion they authored. They compare fixed-response and open-ended items as data sources and identify a theme no closed item offered.
Byproduct: coded dataset, each student's codebook and documented prompt.
Steering the output, and the policy. Students distinguish bias originating in training data from bias installed after training, and locate control at the layer where change is cheap: weights are expensive and frozen; system prompts, tools, and wrappers are editable in seconds.
Students demonstrate the second kind — same user prompt, different system prompts, documented difference in output — and read published system prompts as primary source documents.
Students build a timeline of a documented incident in which an instruction change preceded a behavior change, using public commit history, and evaluate the competing explanations offered by the company and its critics.
Students draft an individual AI-use policy, synthesize the drafts with an LLM, locate their own position in the synthesis or find it absent, and revise the synthesis prompt to preserve dissent. The class adopts the resulting document.
Students distinguish uses of AI that build capability from uses that substitute for it, and state what performance under no-AI conditions requires.
Byproduct: the class AI policy; hands off to Module 2.
Module 2 — Reading the Machine
Sep 9 – Oct 2 · 18 daysModule 3 — Authoring the Machine
Oct 5 – Nov 20 · 31 daysSpecify → generate → judge as the craft: how to command what you can't hand-build
(decomposition)
Read & judge the code AI generates — is it right? make it better
↳ Lesson plan: Sketch, Spec, Swap(tracing → MCQ)
Build with objects you direct — command capability you didn't author
(object model)
How it travels — what actually happens when you send a request
(networks)
Build something that works with text and data
(String reasoning, arithmetic)
Redesign the tools you've been using all semester
(human-centered design)
Capstone build — a genuinely useful application · AI Essentials wrap
Module 4 — How Programs Work
Nov 30 – Dec 15 · 11 days · Fall LEAP Dec 1–18, Q2 benchmark Dec 3–10Short, fragmented, and deliberately so. Reading work survives interruption in a way build work does not — a student pulled for testing can rejoin a trace; they cannot rejoin a group build. This module is the last term in the sequence: data → machine → authoring → the program itself. What you spent a semester directing becomes legible.
The seeding is real but silent. Fall's job is students who leave fascinated by what they can build and unafraid of complexity — the best on-ramp there is. Every activity here must clear the Identity Charter's §4.1 gates.
Spine check
- §AI-building is the course; the CED seeding (object model, tracing, data ethics) is a silent byproduct that must never become a week's real purpose — Identity Charter §4.1.
- §Late arrivals: individualized catch-up via the free quarry + the daily SDD loop; never slow the cohort.