How this was built

Skill: course-build → write-lesson (write-example-sequence, design-check-sequence, write-mcqs), write-video-script → video-brief, review-lesson, write-like-james @ vault 658e971

The shape of the topic

The format ladder is the one James set for the grade 3 to 8 courses: a topic intro that teaches nothing, then per lesson a video of at most 75 seconds (150 spoken words) or an article, ending "Okay, now your turn" with one easy question, then a quiz of about five to eight quick items, then a summary video with three questions, then mixed practice with the terms by recall. In a full course a bank of fifty tagged items feeds the mastery test after this.

The six lessons and their shapes

#	Lesson	One idea	Sequence shape	Quiz type
1	A criterion says the job	A criterion is a job the design must do, that anyone can test	categorical: non-example, three examples, near-miss, rule	fluency, case + yes/no, term item last
2	A constraint says the limit	A limit on what you may use: materials, time, money, space	categorical: a criterion as the non-example at each end	fluency sort, criterion / constraint, term item last
3	A test does what the criterion says	Then you watch what happens	categorical: three pencils against ten; looking against bumping	fluency, does the test check it? yes/no
4	The same test for every design	Then the only difference is the design	one pair of holders: different bumps, then the same bump	fluency, fair? yes/no
5	Met, not met, or a limit broken	Solved when every criterion is met and no constraint is broken; if one job fails or one limit is broken, not yet	three-way sort on the two holders	fluency sort, four result words; two yes/no on the rule
6	Change the part that failed, then test again	Never change the test to make the design pass	worked improvement with the non-example of changing the test	fluency, which part? and yes/no

Rules applied that you can check on the pages

Minimal difference. Every non-example differs from its neighbour by the deciding feature only, in the same sentence frame: "That is a criterion." / "That is not a criterion."
The rule last, word for word. The video's rule beat and the article's rule box are the same sentences. The video's closing bullets are those sentences, because the generator speaks the bullets.
Explain, then label. "Anyone can count ten pencils. That is a criterion." before "A criterion is …".
One word for one thing. Three coined terms in the whole topic. "Job" and "limit" stand beside them.
Fresh contexts in the checks. The lessons use a pencil holder. The quizzes use bird feeders, paper bridges, kites, bookmarks, boats, planters, phone stands, marble runs, a seedling shade.
One demand per item, no verdict bundled with a reason. Two or three options for a sort. The reason lives in the explanation shown after the click.
Keys spread. Attached questions, in video order: A B C B B A A C B. No four keys in a row alternate in any quiz.
Every sentence a student reads has a subject and a verb; every number carries its unit word; no figurative verb.

Quality gates run

check_brief.py on the Peps batch: clean except three accepted hits recorded in READY.md (the intro has no bullets by design; two rule sentences are 13 words).
The two term items in the mixed practice are recognition items, not typed: a grade 4 child cannot reliably spell "criterion", and the rulebook says such a word is recognised among its sisters, never typed.
Two blind key readers, who saw only shuffled, unmarked options, answered every item cold and agreed with every key; a non-author agent reviewed every item against the question-writing checklist (records/REVIEW_QUESTIONS.md); the Grade 3 build's coordinator ran a rules audit against every rule in force on 2 October (records/RULES_AUDIT_from_g3_session_2026-10-02.md). Every flag was adjudicated in records/ADJUDICATION.md, the rewrites went to a third and a fourth fresh reader, and the eight videos were re-rendered from the corrected scripts.
The Grade 3 build's own mechanical sweeps were run on a mirror of this content in the Grade 3 schema (records/g3schema/): word-level readability, the subject-and-verb sweep, key spread, stem length, script length.

Files

content.py is the single source: scripts, articles, questions. The site, the Peps batch and the blind bundle are built from it.
handoff/CCDEMO/scripts/ the eight video scripts; handoff/CCDEMO/videos/batch_CCDEMO.json the batch sent to the pipeline.
EVALUATION.md the review of Andy's script; records/ the check plan and the blind-read reports.