Three constraints on the purpose-built tools that are load-bearing curriculum decisions rather than scoping ones. Each is the reason a claim above survives contact with a classroom.
Why reading is the right substrate
Reading is the substrate for the verification pipeline, and the correctness and teachability asymmetry that Unit 2 rests on is cleaner on reading than it was on math. Correctness is a key match: free, binary, machine-checkable. Teachability is naming which trap a distractor is, why someone would fall for it, and what in the passage settles it. That needs a rubric and a human. A student can explain why a distractor is tempting without being able to work the problem the item is drawn from, which matters in a required course where nobody should be locked out of the teachability judgment by their math level.
The course reads sources regardless, so the item bank is a byproduct of the reading rather than a project of its own. By December that is fifteen to eighteen worked passages on tools, capabilities, jobs, and the industry, each of which is simultaneously the reading, the ACT practice, the Unit 2 corpus, and the evidence base for the class AI policy.
Every item carries at least one labeled distractor type. The four are: true in the world but not in the passage; true in the passage but answering a different question; half right, wrong attribution; and extreme wording (always, never, all, none, proves). The label lives in the key, not on the student's screen.
MuggsOfMath is retired for the pilot year. Not deferred, retired. Google Forms with a locked column schema does everything the model-free core was specified to do: item loads, answer captured, string comparison against a key, export to the longitudinal table. The tool was solving a problem the existing pipeline already solves. Item codes MB-01 through MB-10 stay on disk, and year two decides whether they come back.
Paper before software, and why independence matters
It runs on paper before it runs anywhere else. U2-02 through U2-05 is groups answering reading items and rotating; U2-06 through U2-10 names what happened and designs the scaled version. Experience before abstraction, the same order as vibecode before spec panels, and paper before the tool in Unit 3. Nothing in U2-01 through U2-10 requires software, which also means Unit 2 opens with a floor under it. Independence is the design detail that matters most: a group reading a passage together produces one reading with four names on it, which is review but not redundancy. Five silent minutes first gives four independent attempts per item, and convergence becomes something the teacher watches rather than infers.
The closed loop — they study from what they build
The corpus is the durable asset, not the tool. Tools get rewritten. Models change, frameworks change, the interface will be rebuilt at least twice. A verified bank of 160-odd items with distractor mappings survives all of it, which is why the schema deserves more care this fall than the tool does. At three-way redundancy across roughly 160 items with 75–80 students, each student authors somewhere between six and eighteen of them, about 4–11% of the bank, so if any of these students later studies from it, the vast majority is fresh to them. Contamination is smaller than it looks.
On reading it is smaller still. The math version had a real contamination question: a student who authored an item and later studied from the bank had seen it. On reading, every item is attached to a passage. A student who re-reads a passage they worked in September is doing the thing the practice is for. Recognizing your own item is review, not contamination. Authoring an item is stronger preparation than practicing it in either version, and recognizing your own item later is a good moment.
Curation — the class drafts, the instructor is the target pass
Curation is a stage in the design, not a repair of it. The corpus the class produces is a draft, and it is cleaned before any of it reaches the tool. That relationship has a name students meet in Unit 1: speculative decoding. A cheap draft model proposes several tokens at once; a stronger target model verifies them in a single pass, keeps the run that matches, and corrects at the first divergence. Here the students are the draft — three of them independently working the same item, cheaply and in parallel — and the instructor is the target pass that accepts or corrects before anything ships. The draft is not the lesser thing in that arrangement. It is what makes the expensive pass affordable, and a draft that is usually right is the whole reason the scheme is worth building.
What the design requires is that the target pass be visible rather than silent. If the bank is trustworthy because the pipeline the students designed worked, the reliability numbers mean something. If it is trustworthy because everything was quietly rewritten afterward, the pipeline is decoration and the numbers measure something that does not determine the outcome — precisely the failure this course exists to name. Made visible it becomes data: of the items that cleared three-way consensus, what fraction still needed substantive correction? Five percent means they built something good and they will know why. Forty percent means their thresholds were wrong and they will want to know where, which is the best possible thing to hand a class that just spent two periods arguing about consensus rules.
What keeps the correction figure honest
Three rules keep that number honest. Key-gate failures are blocking, not a review queue — an item whose solution does not reach the keyed answer never ships to students, full stop, because the stakes here are a wrong solution studied from weeks before a test tied to TOPS eligibility. The two kinds of edit are separated and only one is counted: normalizing notation and tightening prose are invisible and need no logging, while fixing a wrong intermediate value or restructuring reasoning is substantive, and only that category belongs in any reported figure. And the pass is committed in two steps — the generated version, then the cleaned version — so the git history is the audit trail and the precision analysis generates from the diffs rather than from anyone's memory.
Why a weak model is the right instrument
A frontier model would ruin both MuggsOfPrompts labs. It produces good output from a vague system prompt, and a student who writes something careless and gets something decent concludes prompt engineering is vibes. A small model's weakness is the instrument's sensitivity — vague instructions produce visibly worse output, specific ones visibly better — and flatter probability distributions mean clicking a token shows genuine competition between candidates rather than a column of 0.99s. This has a direct consequence for fallbacks: comparable must mean comparably small, not comparably capable. Falling back to a frontier API does not preserve the lab; it inverts it. The same holds for MuggsOfCode, where the loop depends on a weak spec producing visibly weak output.
Stated versus enforced
Stated versus enforced is the transferable idea, and it comes from the tools themselves. MuggsOfSources has the model emit sentence IDs, so fabrication is structurally impossible rather than discouraged — the guarantee lives in the architecture, not in an instruction the model is asked to follow. MuggsOfCode forbids the AI from writing spec content and enforces it by never routing model output into the spec panels. A student who can ask is this behavior promised, or enforced? has a test that outlasts every tool in this course.
Highest-priority build item. U1-15 through U1-18 (tokenization live, candidate probabilities, why token count disagrees with word count) and U1-29 through U1-31 (the system-prompt battery with a fixed seed) land in the second half of August under the operating-spec ordering. Earlier v18 artifacts listed Aug 19 and Sep 1; the current build-need-by date is Aug 24, aligned to the Week 3 opening under the calendar's day-by-day plan. Minimum viable form: fixed seed, fixed battery, outputs, and the seed displayed. The entire what-happened panel sits behind one toggle and can ship empty.
Not the soonest, but the deepest — sentence IDs, attestation, extraction damage, and quotes-by-selection are U2-11 through U2-19 objectives rather than conveniences. That block has no floor without it.
MuggsOfMath
retired for pilot year
Retired, not deferred. Google Forms with a locked column schema does everything the model-free core was specified to do: item loads, answer captured, string comparison against a key, export to the longitudinal table. The tool was solving a problem the existing pipeline already solves. Unit 2 has moved to ACT-style reading items on passages the class already reads, which turns the item bank into a byproduct of the reading rather than a project of its own. Item codes MB-01 through MB-10 stay on disk; year two decides whether they come back.