Skill: course-build → write-mcqs → review-questions → review-content → write-like-james @ 16bb23f

Grade 4 end-of-grade test: built, checked, cold-read

Relabelled 8 Oct 2026, no redraw: the two Design and build items on each form are Physics items (Topics 11 and 12 are Physics topics). James, 8 October 2026 (reviews/james/RULINGS_2026-10-08.md, q-G5ARCH-1 and -2): "I think fold all the design and build stuff into Physics — even if it's its own topic — same with G4 — I think we can make that a physics topic"; "PP100 — and we'll cover it inside physics." The form modules still carry "strand": "DAB" on those two items and build_tools_check_grade_test.py still plans DAB 2; both are the grade-test build's to relabel.

What happened. Grade 4 now has its end-of-grade test: three fixed forms of 36 questions, A, B and C, no question shared between them. Every form samples all five strands in proportion to how much each teaches. 42 questions are written in the shape of a real state-test item (STAAR Grade 5, Florida Grade 5, Arizona Grade 5) whose standard is a Grade 4 atom. The rest come from the topic PP100 banks and are marked there so the PP100 stops serving them.

What it means for a child. A child who passes has answered the state tests' own Grade 4 content, in the state tests' own shapes, a year before the state tests ask it. The pass mark is 33 of 36.

What you decide. Nothing blocks. Every format default is Grade 3's shape or a default I set, each marked below, and all of them wait for Becky's validation. Say so if you want any of them the other way. The one choice with teeth is the pass mark: 33 (91.7 %, the first score at or above 90 %) or 32 (88.9 %).

Where the quality stands. The checker passes at 0 FAIL. Three cold readers, one per form, who never saw the keys, agreed with all 108 keys. Their flags and the non-author examiner's findings were applied the same morning; the list is below.

The defaults, one line each

36 questions a form. Grade 3's endpoint length. The Grade 3 assessment start file mentions 24 or 30 only as candidates for a blueprint not yet written. Pending Becky.
Four-option questions only. Grade 3's shape: only item types the platform renders and the child has met. Pending Becky.
Pass at 33 of 36. The first score at or above the 90 % convention. Pending Becky.
No strand at zero. Grade 3's floor. With three misses allowed it never bites now that Design and build sits inside Physics (12 questions); it bit only when Design and build was a strand of two. Pending Becky.
Three attempts: A, then B, then C, with teaching in between. Below 26: back to the lessons of every missed topic. 26 to 32: redo those topics' PP100s, then the next form. A third fail: the teacher decides. Grade 3's shape, pending Becky.
Untimed, about 40 minutes for a ready child. Pending Becky.
Strands by atom count: Physics 12 (10 from Topics 1-10 and 2 from Topics 11-12, the design-and-build topics) · Matter 3 · Earth and Space 15 · Life 6. Design and build rounded up from 1.6 to 2 so both its topics sit on every form; Earth and Space gave up the one. Pending Becky.
Depth: about 7 recall, 21 apply, 8 transfer a form. Every hard question says in one line why it is hard. Pending Becky.
Two written one-sentence questions a form are written but held back until the AI marker has been checked against a human marker on 50 children's sentences. Grade 3's shape.
Drawn bank questions are withdrawn from the PP100. The mark is in the bank; the platform side still has to be built. Pending you and Becky.

The seeds: what the state tests gave us

Found and read: the STAAR Grade 5 Spring 2025 answer key and item rationales (every item described); the STAAR 2023 Grade 5 practice test in full; the STAAR 2024 practice key (standards only, no item text); Florida's Grade 5 sample test with its key and rationales; Florida's Grade 5 item specifications (one sample item per benchmark, including the Grade 4 ones); Arizona's AzSCI Grade 5 sample test scoring guide (standards, phenomena and rubrics; the stems are pictures).
Used: 42 released items on Grade 4 atoms became one question each, in the released shape, with our scene and words; 2 more shapes already existed in the banks and were drawn as they stood. The identifier sits on every seeded question.
Not found: the STAAR 2025 and 2024 forms as text (they are online-only now), STAAR Grade 5 forms before 2023, and Arizona's operational items (never released). David's own harvest of released Grade 5 items is on his machine.

What the readers found and what changed

Three cold readers agreed with every key, 108 of 108. A second reader re-read every item that then changed.
Two questions had a second defensible answer and were replaced (a seed pot in a garage that is dark as well as cold; a still turbine where the electricity is "missing" too); one erosion distractor (a pebble worn smooth in a stream) was replaced; one bank item on fog "all day" made "none at all" arguable and was swapped.
Wording was fixed on six seeded questions; six bank questions were swapped for cleaner siblings from the same bank; two held written questions moved to other atoms so no question on the form hands them their answer.
Judged and left: the banks' own "which two parts of Earth act here?" wording and their "none of these" option (the course's taught forms); the criterion-or-constraint items a cold reader found open (they test the coined terms the lessons teach); the crust's "about twenty miles" (the lesson's own number).
The examiner's pass (a second non-author read against the question-writing checklist) found two must-fixes, both applied: one explanation pointed at the wrong option after the keys were spread, and one stem asked "which sentence shows he is wrong", the shape you barred. Of its 45 smaller findings, 41 were applied (plainer asks, distractors no child would pick replaced, five "hard" tags lowered to medium, British spellings out) and 10 left with a reason beside each.

What happens next

Becky's validation of the defaults; your word on 33 or 32 and on withdrawing drawn items from the PP100.
The grader-vs-human check, then the six held written questions join the forms.
The three forms are now the record (grade4/grade_test/form_A.py, form_B.py, form_C.py); the full README sits beside them.