Skill: course-build → write-mcqs → review-questions (examiner) → review-lesson (cold reads) → write-like-james @ 465e657

Grade 3 Science end-of-strand tests: four tests, three forms each, built 8 October 2026

In short. Your word tonight was "build the Grade 3 end-of-strand tests … fresh items … the same kind of standard as the actual exam". Twelve forms now exist: Physics, Matter, Earth and Space, Life, each as Forms A, B and C of 30 questions. Every one of the 360 questions is new (no bank, quiz, practice or video question re-used; a word-overlap sweep plus a non-author examiner's eye for re-dressed lesson scenes), and every key was agreed by a cold reader who had not seen it, twice where an item changed. Nothing is live, nothing is in the Drive package, and every format default below is still "pending Becky".

One question for you (and Becky)

How hard should a strand test be when the pass mark is 27 of 30?

Your words tonight set the pitch: the same standard as the real exam, which is about 40% judge-a-claim or read-data items. Tagged honestly, the forms carry 9 to 14 hard items each (see the table). The 30 September format proposal, written for a readiness gate, recommended 6 to 8 hard items at a pass mark of 27; the Matter examiner's arithmetic (David's own targets of 92% on medium and 84% on hard items for a child who has mastered the course) puts a course-master's expected score just under 27 with 12 hard items and just over it with 8.

• Recommendation: keep the exam's pitch and lower the pass mark to 26 of 30 on these forms, so a child who knows the course passes about three times in four rather than one in two. Becky validates the figure.
• Choices: 27 with the current hard counts · 26 with the current hard counts · 27 and cut each form to 8 hard items (swap the rest for apply items).

Nothing else needs you. Three matters go to other people: Scroll must confirm it shows a small table inside a question and a shared table above each item of a cluster (every data display is one); the released-exam habit of a picture on most items waits for the picture stage; and the spec's 15-word sentence limit pulls against our own checker's three-sentence rule on medium items (a note for Becky).

The forms at a glance

Key letters sit at 20 to 30% each on every form, with no run of three. Items sit in topic order, in proportion to each topic's atoms (every Physics topic has two; Matter 5/5/4/4/6/6; Earth and Space and Life floor two with the atom-rich topics taking more). Two one-sentence written items a form are written and held out of the 30.

The format (the 30 September proposal's recommended shape; every line pending Becky)

• 30 items a form, untimed, about 30 minutes for a ready child; four-option multiple choice, with a typed number where the exam would have the child type one (a reading off a drawn instrument, a difference, a count).
• Pass at 27 of 30 (the question above). No topic may score zero: a topic at zero routes the child to that topic's lessons whatever the total.
• Three fixed forms, taken A then B then C with teaching between: below 21, back to the lessons of every topic with a missed item; 21 to 26, that topic's PP100 first; a third fail, the teacher decides.
• Clusters of two or three items on one scene or data table, as the Alpha exam and the state tests do; each item still names its scene so it stands alone.

How they were built and checked

1. Four authors, one per strand, wrote every item fresh under the write-mcqs skill, with a design-move inventory (each item's move, scene kind and nearest earlier item) and the Grade 3 language rules. A freshness sweep (check_fresh.py) compared every stem, option and shared table against all 5,207 earlier Grade 3 items and the released state texts: 0 hits on every form.
2. One cold reader per form answered the 30 items with no key: 360 of 360 keys agreed. Their construction flags are in records/READS_LOG_2026-10-08.md.
3. A non-author examiner per strand judged every item against the Alpha Grade 3 spec, the released Grade 5 items' standard and the course's rules: no wrong keys; 32, 32, 14 and 4 must-fix rows (Earth and Space, Matter, Life, Physics), including 15 items replaced because they were a lesson scene in new words or not exam items, four keys false in the real world, and over-tagged depth.
4. Each author applied one consolidated fix pass; every changed item was re-bundled and read again cold: every key agreed. Final polish (plain words, stand-alone stems) changed no key.
5. Checkers at the end: topic checker 0 FAIL on all twelve forms; key-tell checker 0 flagged; freshness 0 hits.

Where things are

• Forms: grade3/strand_tests/G3S_<PHY|MAT|EAS|LIF>_form<A|B|C>.py (schema as Grade 4's, plus IS_TEST, stimulus for a cluster's shared table, cluster tags); drawn instruments in figures_MAT.py and figures_EAS.py.
• Records: grade3/strand_tests/records/BUILD_<STRAND>_2026-10-08.md (James's words, defaults, strata, inventory, every fix), EXAMINER_<STRAND>_2026-10-08.md, READS_LOG_2026-10-08.md; blind bundles, sequestered keys and readers' answers under blind/2026-10-08/.
• Not done, on your word: the practice end-of-grade test (36 fresh items, never the protected exam items) is a "maybe" for later; no real end-of-year test.

What happens next

Your ruling on the pass mark; Becky's validation of the format; Scroll's answer on tables; then the forms are exported to the platform the way the PP100 banks were. Until then they live here only.