Skill: course-build → write-lesson, write-video-script, write-mcqs, design-check-sequence, sentence-atomise, write-explanation, review-figures @ b0fa825

Skills in full: course-build → write-lesson (L01 a fact with its mechanism — deeper is older because each new layer lands on top; L02 a routine — count up from the bottom; L03 a fact — a fossil is as old as its layer, the sinking belief confronted by name; L04 the capstone routine — read each layer from the bottom up, then say the changes in order) → write-example-sequence (considered; no categorical concept, so no example run), design-check-sequence, sentence-atomise, write-explanation, write-mcqs; write-video-script (the intro, four lesson scripts, the summary; the elementary micro-lesson anatomy); review-figures (author self-check on sixteen renders; a non-author pass is owed) @ b0fa825

Grade 4 Topic 21 — How rock layers tell the time — build record (v1, 2 October 2026)

What this is, in plain words

Four short lessons, one per atom, in one class: Miss Ruiz's, with Jordan, Julia and Karl, on a trip that stops at a striped road cut in Texas and later at a hillside a river has cut into near their school. First the child learns why the bottom stripe formed first: each new layer of pieces can only land on top of the layers already there, so in layers that still lie flat, deeper is older. Then the routine: to put layers in time order, start at the bottom and count up, 1 for the layer that formed first. Then a fossil: it was buried as its layer's pieces settled on it, so it is about as old as its layer, and a fossil in a deeper layer is older than one above — and no, a heavy shell did not sink down through solid rock. Last, the capstone: read each layer from the bottom up with Grade 3's fossil rule (what the layer holds → where that living thing lives today → what the place was like), and the changes, in that order, are the place's story: sea, then sandy shore, then forest. Built with no blueprint, under the Grade 4 standing brief; the structural decisions are below. Nothing is sent and nothing is committed.

Structural decisions (in place of a blueprint) — one line each, with the reason

• Four lessons, one atom each, in the register's order (035 → 036 → 037 → 038): every prerequisite edge points backwards; 038 is the capstone.
• One running scene, the Texas road cut, for L01–L03; a second cut (a river-cut hillside near the school) for L04: the road cut's four stripes carry the fact, the count and the two fossils, but a sea → shore → forest story needs shells, sand and a soil with roots, which the road cut does not have; a second place is more honest than re-drawing the first.
• L01 is the FACT form with its mechanism inside it (the register's JM-P6 exception): "deeper is older" is unusable at this grade without "because each new layer lands on top", so the landing strip (one layer, then a second on top, then a third, then the dry cliff) IS the lesson; no example run, because the only non-example (tipped or broken layers) is the sibling case grades 6-8 own.
• The jar from Topic 18 and the river-at-the-sea from Topic 20 are recalled, not re-taught: each by one line (Topic 18's and Topic 20's lexicon sentences word for word) and one recall item; the jar is Topic 18's own drawing, imported.
• "Layers that still lie flat" is said in the rule and shown in every diagram: the simplification is named ("Tipped or broken layers wait for a later grade" was cut for length; the article keeps "This rule is for layers that still lie flat, the way they landed"), and no diagram in the topic draws a tilted band.
• Deep time stays Grade 3's "a very, very long time", with one number: "hundreds or even thousands of years" for one layer: the brief's misconception (layers form in days) needs a number a child can hold; a flood layer can form faster, so the article says "one layer takes a very, very long time to form: hundreds or even thousands of years" and the record names the simplification here.
• L02 is a ROUTINE worked on five layers then run on three, with the number chips appearing one by one from the bottom: Julia's drawing adds a thin soil stripe on top so the count reaches five; the time words (formed first = oldest = bottom; formed last = youngest = top) are tied together in their own short section because the items use both.
• L03's mechanism is the burial strip (shell on the sea floor → pieces settle → buried in its layer → more layers on top), and the register's sinking belief gets its own card ("Does a heavy shell sink down through solid rock?" — NO: ROCK IS SOLID): a verdict mark names its question.
• L04 quotes Grade 3's two sentences word for word and then says the sameness out loud ("read one layer" → "read every layer, from the bottom up"): the register's cross note asks for it; the story table (LAYER | WHAT IT HOLDS | WHAT THE PLACE WAS LIKE) is the routine's record, filled from the bottom row up; the second cliff is the first's reverse (forest → shore → sea) so the same readings in the other order give the other story.
• No term is coined: "layer", "fossil", "sedimentary rock" and "deposition" are earlier topics' words; "oldest / youngest" are everyday; strata, superposition and relative age are grades 6-8. So there are no term-recall items; the hard tier is typed where typing produces (a position number counted from a drawing, the count of layers above a lettered layer, the word "layer") and chained MCQs elsewhere. Question for James below.
• Every layer diagram is drawn by code (cliff()), with one drawing register: five layer colours (tan sand, gray mud, cream shell rock, red-brown sand, dark brown soil), Grade 3 Topic 27's fossil icons imported and drawn small in a band's darker tone, number chips on the left (1 at the bottom), FORMED FIRST / FORMED LAST chips on teaching figures only, letters J–P on item figures; the headache figure is the teaching figure with the number layer off.
• Real cliffs and fossils are photo placeholders, the class is a cartoon placeholder: two photos of real striped rock (a road cut; the Grand Canyon for the intro), four fossil photos imported from Grade 3 Topic 27, four cartoons.
• All four lessons are "video and article": each carries a drawing that moves (layers landing; chips appearing; a shell being buried; a table filling), which the video shows and the article shows as a strip; none could do without.
• The intro is Jordan at the Grand Canyon rim with his dad, a sea shell in a rock high above any sea: one scene, one felt question ("which stripe came first, and how a sea shell got up here"), one question mark in the narration, the map line, the ready line and the poll.

Questions for James (the build does not wait)

• Should this topic coin a term, or stay term-free as built? Context: the four atoms name no term; Grade 3 coined "fossil", Topic 18 "sedimentary rock" and "sediment", Topic 20 "deposition". Candidates would be "rock layer" (everyday already) or "time order" (a phrase, not a term). Recommendation: stay term-free; the bank's hard tier is typed numbers and a chained MCQ, which test the skill. Built term-free; RECALL_BY_MCQ is not needed.
• Is "hundreds or even thousands of years" for one layer the right number to say, or should deep time stay as "a very, very long time" only? Context: the brief names "children think layers form in days"; the atom gives no number. Recommendation: keep the one number, said once in L01 and tested once (Karl's claim, Yes). Built with it.
• Should L01 and L04 be split at a check, at 344 and 349 article words? Context: the brief's ceiling is 300, with WARN above it acceptable when a subtitle or figure sits every ≤ 8 lines (true here: 5 subtitles and 6–7 figures each, never more than 4 idea lines without one) and a reason is recorded. Reason: L01 carries the recall of two earlier topics (two lines), the landing strip, the two beliefs confronted by name and the scope line; L04 carries Grade 3's two quoted sentences, the worked case, its table and the reverse case. Recommendation: keep each as one article. Built as one each.

Counts

4 lessons · scripts 150 / 149 / 150 / 150 spoken words by the checker's count (146 / 145 / 146 / 146 by the batch tool's; cap 150) · intro 104 words, 1 question mark · summary 150 words (cap 175) · article student text (idea + rule + subtitle lines) 344 / 256 / 281 / 349 (Grade 3 median 308; brief's ceiling 300 — L01 and L04 answered above) · idea lines 19 / 14 / 16 / 20 (longest 22 / 21 / 22 / 22 words) · subtitles 5 / 4 / 5 / 5, figures 6 / 6 / 7 / 7 per article (never more than 4 idea lines without one) · quizzes 10 / 10 / 10 / 11 items (one reorder in L02) · attached questions 4 + 3 (summary), key string in video order C C A B A B A (A 3 of 7) · practice 12 (9 MCQ, 1 reorder, 1 typed number, 1 written; 4 retrievals: Topic 18, Topic 20, Grade 3 Topic 27 ×2 — the fourth folded into a drawn story item) · PP100 bank 51 items: 46 four-option MCQs, 2 reorder, 2 typed numbers, 1 typed word; EAR4-035 13 (8 easy, 4 medium, 1 hard) · EAR4-036 13 (6 / 5 / 2) · EAR4-037 12 (3 / 6 / 3) · EAR4-038 13 (2 / 5 / 6); keys on the four-option MCQs A 9 · B 13 · C 13 · D 11 · figures 63 (53 SVG, 6 photo_ph (4 imported from Grade 3 Topic 27), 4 cartoon_ph).

The per-lesson table

design-check-sequence records

`

Skill: design-check-sequence @ b0fa825

Part: L01 (one part)

Type: (a) — the student reads a drawn or described cliff for the layer that formed first or last: one step, one label; plus the claim items.

Plan: 10 items; 2 recall (3 options) → 2 drawn fluency (letters; 3 and 4 options) → 2 fluency (3 options) → 3 claims yes / no (n n y) → 1 apply (3 options). Stems ≤ 25 words.

Belongs in practice/test instead: the reason items (PP type 2), the two-relation chain (PP-40), thickness-is-not-age (PP-22).

Part: L02 (one part)

Type: (b) — a routine (start at the bottom, count up) whose steps each have a right answer; the student can be right on number 1 and wrong on the next.

Plan: 10 items; 1 recall (2 options) → 3 scaffolded on a SECOND drawn case J K L M (number 1; right after the bottom; last) → 1 reorder on a THIRD drawn case (K L M N P, letters out of order) → 1 drawn youngest (J L N) → 2 bare, no drawing (the top layer's number; the oldest) → 1 claim yes / no → 1 apply. Four-option keys 3 2 0; three-option 1 2 0 0; two-option 0 1.

Belongs in practice/test instead: typed position numbers (PP type 6), the count-down item (PP-27), two cliffs of different sizes (PP-28).

Part: L03 (one part)

Type: (a) — the student places a fossil's age by its layer: one step; plus the claim items.

Plan: 10 items; 1 recall (3 options) → 1 when-buried (2 options) → 2 drawn older / younger (2 options; cliffs J K L M and K N P) → 1 same-layer (3 options) → 3 claims yes / no (n y n) → 1 apply (2 options) → 1 transfer (3 options). Two-option key string 0 0 1 1 0 1 1 (no four alternate).

Belongs in practice/test instead: the reason items (PP type 8), the opposite-ends chain (PP-43), the typed "layer" (PP-45).

Part: L04 (one part)

Type: (b) — a routine (read each layer from the bottom up, then say the changes in order) with a right answer at every step.

Plan: 11 items; 1 recall (3 options) → 4 scaffolded on a SECOND drawn case J K L (what the place was like when J formed; when L formed; which layer tells the first part; which sentence tells the story) → 2 bare drawn on a THIRD and FOURTH case (M N: how did it change; M N P: which story) → 1 bare, no drawing (2 options) → 3 claims yes / no (y n n). Three-option keys 0 1 1 0 2 1 2; two-option 0 0 1 1.

Belongs in practice/test instead: the return-to-the-sea stories (PP-47, PP-50), two cliffs compared (PP-51), the written explanation (PR-12).

`

The key strings

• Attached questions, video order: C C A B A B A (A 3 of 7, under 60 %).
• L01 quiz: 3-option 1 1 2 0 0 0; 4-option 0; 2-option 1 1 0. L02 quiz: 2-option 0 1; 4-option 3 2 0; 3-option 1 2 0 0. L03 quiz: 3-option 0 2 1; 2-option 0 0 1 1 0 1 1. L04 quiz: 3-option 0 1 1 0 2 1 2; 2-option 0 0 1 1.
• Practice: 3-option 1 2 2 1 0 2 1; 2-option 0 1.
• Bank four-option MCQs: A 9 · B 13 · C 13 · D 11 (46 items); reorder keys N J P K and J N L; typed 4, 2, layer.

The example sequences and worked cases, written out

• L01 (fact form, examples only): the road cut's four stripes → each stripe is sedimentary rock (Topic 18's sentence; the jar) → a sea floor: a river drops sand and mud in one layer → a second layer can only land on top → a third on top of that → hundreds or thousands of years each → today, rock in a cliff, numbered 1–4 from the bottom → Julia's top-is-oldest (No) and the all-at-once belief (No) → the scope line → rule. Non-example considered and rejected: tipped or broken layers (the sibling case, grades 6-8).
• L02 (routine, worked then second case): Julia's five stripes → 1 on the bottom → 2, 3, 4 up → 5 on the soil → a three-layer cliff: tan 1, gray 2, cream 3 → formed first = oldest = bottom; formed last = youngest = top → rule.
• L03 (fact form): the shell at the bottom and the leaf at the top → Grade 3's fossil sentence → a shell on the sea floor is buried by settling pieces that become its layer → as old as its layer; later layers younger → Jordan's sinking (No: rock is solid) → shells in layer 1, leaf prints in layer 4 → the shell is older → rule.
• L04 (routine, worked then its reverse): shells / sand / soil with roots → Grade 3's two sentences → read layer 1 (under the sea), 2 (a sandy shore), 3 (a forest) into the table → the story: sea, then shore, then forest → the reverse cliff (fern prints / sand / shells): forest, then shore, then the sea came in → rule.

The rules applied

Misconception rows (the catalogue is silent on rocks; the register's SD-3 rows and the brief's three beliefs, with the literature named)

Checker: 0 FAIL, 8 WARN (python3 build_tools_check_topic.py grade4/topic21_build/content/G4T21.py)

Three S13 WARNs and five S36 WARNs. Answered one by one:

python3 grade4/check_g4_articles.py grade4/topic21_build/content/G4T21.py → 0 FAIL (articles 344 / 256 / 281 / 349 with the two WARNs answered above; scripts 150 / 149 / 150 / 150; intro 104; summary 150, opener ok).

Names run (python3 build_tools_names.py Ruiz Jordan Julia Karl → exit 0, all four free)

Miss Ruiz; Jordan, Julia, Karl — the Topic 21 row of grade4/NAMES_POOL.md. No other name appears in a student string ("dad" in the intro is unnamed).

Bank freshness (re-performed; the scene-noun grep)

Scene nouns of every lesson, quiz, attached and summary item: road cut, highway, Texas, the four stripes, notebook, Julia's drawing, the jar, the sea floor and the river, the Grand Canyon and Jordan's dad (intro), the river-cut hillside near the school, the shell and the leaf at the road cut, a canyon wall in Arizona (L01 q10), a stream and a pond (L01 q2), a geologist digging through five layers (L03 q9), a dry hill (L04 q1). Grep of the bank's stems and options (comments excluded): road cut 0 · highway 0 · Texas 0 · notebook 0 · jar 0 · canyon 0 · dad 0 · school 0 · river 0 · hillside 0 · stream 0 · pond 0 · hill 0 (the Georgia item says "a hill behind a new store" — its scene is the cut, not a hill; kept). Re-scened on the first pass: "a hill behind a school in Georgia" → "behind a new store" (the school is L04's scene); "a hillside in Utah" → "a mountainside"; "a stream bank in Missouri" → "a cut beside a railroad"; L01 q1's "bottom of a lake" → "the sea floor" and L01 q10's "quarry wall in Indiana" → "canyon wall in Arizona" so that lake and quarry are bank-only. The fossil kinds (shell, fish, fern, leaf, footprint, bone, roots) and the three readings (sea, shore, forest) are the topic's drawing register, not scenes, and appear on both sides by design. Bank scenes: a quarry in Indiana, a sea cliff in California, a swimming-pool hole in Florida, a lakeside cliff in Michigan, a mine wall in Colorado, a cut behind a new store in Georgia, a railroad cut in Missouri, a cave wall in Kentucky, an island cliff in Maine, a ditch in Kansas, a mountainside in Utah, cliffs in Alabama and Nevada, a lake floor in a flood, two cliffs side by side.

Blind bundle

grade4/topic21_build/blind/2026-10-02/BLIND_BUNDLE_G4T21.md (102 MCQ, 4 typed (3 numeric), 1 written, 4 reorder; keys sequestered in KEYS_SEQUESTERED_G4T21.json). Every [Drawing: …] line is data ("cliff with 4 flat layers, top to bottom: J tan sand; K cream shell rock; L gray mud; M red-brown sand"); grep for oldest / youngest / older / younger / first / last / order → 0.

Video batch (built, checked, NOT sent)

python3 build_tools_video_batch.py grade4/topic21_build/content/G4T21.py → handoff/G4T21/videos/batch_G4T21.json (6 videos: intro 104 words, L01–L04 146 / 145 / 146 / 146, summary 146) and READY.md; check_brief.py: L01–L04 and the summary ✓, the intro's one standing hit (no closing bullets — an intro has no rule slide). grep -ci strand → 0. Nothing has been sent; James's word is needed for a Grade 4 batch.

Sentence-atomise self-check (one row per lesson; every paragraph read; S1 / S38 clean)

Figures (63; author's review-figures self-check on sixteen renders, a non-author's pass owed)

Rendered with headless Chrome at 1000 px and opened: cliff_numbered, cliff_five_headache, cliff_five_numbered, cliff_five_step1, cliff_story_headache, landing_strip, burial_strip, no_sinking_card, story_table_filled, story_table_two, q_cliff_fossils_a, q_cliff_five_b, q_cliff_story_b, pp_cliff_8, pp_cliff_14, pp_cliff_16, plus jar_recall, deposition_recall and class_roadcut_cartoon (placeholder). Fixed on the render: the burial strip's chip words overflowed their panels (shortened to ON THE SEA FLOOR · PIECES SETTLE · BURIED IN ITS LAYER · MORE LAYERS, number chips moved clear) and its sea floor was the same cream as the shell's layer (now tan under cream, so the shell's own layer stands out); the root icon read as an upward arrow (now a stem from the band's top forking into three tapering roots). Every band is the same width and never tilted; number chips sit 1 at the bottom; item cliffs carry letters J–P and no chips; the table drawer's guard asserts every cell fits its column. Photo and cartoon placeholders render as the pale box with grey text; IMAGE_BRIEFS carry the 10 picture briefs (David's pack, after James's sign-off on the questions above).

Judgement calls, recorded

1. No term coined. Question for James above.
2. One number for deep time ("hundreds or even thousands of years" for one layer), the simplification named here: a flood or a landslide can lay a layer in a day; most layers take far longer; the child needs one honest order of size against "a few days". Question for James above.
3. The sea "moves away" and "comes in", never "pulls back" (the atom's own example wording), because a young reader takes "pulled back" literally.
4. Grade 3's fossil sentence is spoken in L03's article but not its script, for the 150-word budget; the recall item carries it after either.
5. L02's rule is two sentences and L04's is two sentences; the lexicon lists eight sentences, two per lesson, and the summary speaks all eight (150 words).
6. The typed "sea" item was re-authored when S30 refused "sea" as an open key: it became the Alabama / Nevada comparison (PP-51), hard because two stories are compared.
7. Roots as a fossil kind. Tree roots in a soil band stand for a buried forest floor (the atom's own example: "forest-soil layers"); they are drawn from the band's top edge downward so they read as roots, not an arrow.
8. L04's second scene is a river-cut hillside near the school, not the road cut, because the road cut's stripes (shells, mud, sand, red sand, soil) do not tell a sea → shore → forest story without inventing fossils in them.
9. Written to the topic folder and the tools' own output paths only: the page to /tmp/g4t21.html, the figures to /tmp/g4t21_figs/, the batch to handoff/G4T21/videos/ (the tool's path; NOT sent), the bundle to grade4/topic21_build/blind/2026-10-02/. Nothing committed.

Owed

• A non-author review-figures pass and review-questions pass; the blind read of the bundle (keys unverified until a cold reader agrees).
• The picture pass (10 placeholders) after James answers the questions above.
• James's word before the batch is sent (Grade 4's batch rule).
• Topic 20's content is not yet written (only its figures.py exists); the deposition sentence recalled here is Topic 20's rule_deposition card word for word — if Topic 20's lexicon changes, L01 q2's fb / why and RECALL_DEP follow it.
• The S36 additions list: if James accepts "stripe", "fern", "geologist" and "quarry", they go to reference/grade3_word_additions.txt (not written here — outside this topic's folder).

v2 — reviews applied (2 October 2026)

Skill: course-build → write-lesson (article lines), write-video-script (the intro beat), write-mcqs (every item change), write-explanation (fb / why), sentence-atomise (every new sentence), review-figures (the non-author review applied) @ efd8fba

In plain words. Every key held; the drawings did not. cliff() stacks its layers bottom → top, and ten item drawers had been handed their letters top → bottom, so a stem said "fish in J" and the picture put leaf prints in J. The fix is not ten flipped lists but one list per lettered item: figures.ITEM_CLIFFS holds each item cliff once, top → bottom, as (letter, colour, fossil) triples, and words() hands that same entry to the stem, the options, the key (derived: the bottom letter's option index, the bottom-up order, the lower fossil), the fb, the why and the drawing. A guard (check_lettered_items.py) parses every stem, option, fb and why for letter claims and checks them against the figure data: 34 lettered items, 34 match. Around it: the road cut is now one cliff across L01–L03; the typed "layer" item is cut; two summary questions, L04 q9 and eleven bank items have fresh scenes or cleaner options; the FORMED FIRST / LAST chips sit inside their frames; the no-sinking card has a layer under the shell; the landing strip's forming layers show their pieces settling; the intro unpacks "tell the time". Nothing is sent; nothing is committed.

The direction convention. cliff(layers, letters) reads BOTH lists BOTTOM → TOP (layers[0] is the ground band). ITEM_CLIFFS entries are written TOP → BOTTOM, as the eye reads a cliff and as the blind text lists it. item_cliff() is the one place that reverses. An order item's items key is stored FIRST → LAST (bottom → top), the direction every reorder stem asks; both bank reorders (PP-24, PP-25) and the two lesson-side ones (L02 q5, PR-5) ask first to last.

Finding → change

Questions for James — built as recommended

1. Term-free — yes: strata and superposition are grades 6–8's; the one typed word item (Topic 18's "layer") is cut.
2. "Hundreds or even thousands of years" for one layer — kept, said once in L01 and tested once (L01 q9, Yes; PP-8's options now one span each).
3. L01 and L04 whole — yes: a subtitle or figure every ≤ 3 idea lines; L01 is 351 article words (the scope line grew it by one sentence; the jar was a figure, not words), L04 349 — both WARNs answered here as before.

Counts and key strings (v2)

4 lessons · scripts 150 / 148 / 150 / 150 (batch tool 146 / 145 / 146 / 146) · intro 114 words, 1 question mark · summary 150 · articles 351 / 255 / 281 / 349 (idea lines 20 / 14 / 16 / 20; figures 5 / 6 / 7 / 7) · quizzes 10 / 10 / 10 / 11 · attached questions C C A B A B A (A 3 of 7) · practice 12 · bank 50: 46 four-option MCQs, 2 reorder, 2 typed numbers; EAR4-035 13 (8 / 4 / 1) · EAR4-036 13 (6 / 5 / 2) · EAR4-037 11 (3 / 6 / 2) · EAR4-038 13 (2 / 5 / 6); MCQ keys A 9 · B 14 · C 12 · D 11; reorder keys N J P K and L N J; typed 4, 2 · figures 55 (48 SVG, 3 photo, 4 cartoon).

Quiz key strings — L01 3-option 1 1 2 0 0 0, 4-option 0, 2-option 1 1 0 · L02 2-option 0 1, 4-option 3 2 0, 3-option 1 2 0 0 · L03 3-option 0 2 1, 2-option 0 0 1 1 0 1 1 · L04 3-option 0 1 1 0 2 1 2, 2-option 0 0 1 1 · practice 3-option 1 2 2 1 0 2 1, 2-option 0 1. Unchanged from v1 throughout.

Checker results

• build_tools_check_topic.py → 0 FAIL, 8 WARN (the same three S13 lines — both named fossils must be in the stem — and five S36 words, answered in v1).
• grade4/check_g4_articles.py → 0 FAIL (L01 351 and L04 349 WARN, answered above).
• build_tools_names.py Ruiz Jordan Julia Karl → exit 0, all free.
• check_svg_ids.py content/figures.py → PASS (55 figures, 0 ids, 0 shared).
• build_tools_build_topic_site.py → /tmp/g4t21_v2.html; screenshotted at 900 px and read: L01–L04 article routes, L02 and L04 quiz routes, the PP100 route (/tmp/g4t21_v2_*.png). Every item cliff's letters sit beside the bands the stem names; the road cut's leaf prints are in the top red-brown band on L03; the no-sinking card has tan under the cream.
• Every figure rendered (python3 content/figures.py /tmp/g4t21_figs_v2), laid out as two contact sheets (/tmp/g4t21_items_v2.png, /tmp/g4t21_teach_v2.png) and read band by band against the stems: all 28 item cliffs agree; the chips are inside their frames on the three numbered cliffs; LATER and LATER STILL show settling dots.
• check_lettered_items.py content/G4T21.py → SUMMARY: 34 lettered items checked, 34 match their drawing, 0 mismatch (it prints each stem's letter → contents pairs beside the blind text's).
• build_tools_video_batch.py → 6 videos; check_brief.py → L01–L04 and the summary ✓, the intro's one standing hit (no closing bullets); grep -ci strand → 0. Not sent.
• Blind bundle blind/2026-10-02-v2/ (102 MCQ, 3 numeric, 1 FRQ, 4 order; 110 items); its 34 [Drawing: lines grepped for oldest | youngest | older | younger | first | last | order → 0. The four order keys stored bottom → top: L02-05 M K P L N · PR-05 K N J M · PP-24 N J P K · PP-25 L N J.
• v1 → v2 key comparison (key text by item id, PP-46..51 → 45..50): 89 unchanged; 14 changed, all deliberate (L04-07 options trimmed; PP-7, 8, 12, 13, 19, 23, 25, 30, 32, 39 rewrites; PR-7, PP-26, PP-41 numeric, same values).

Items whose stem, options, key, model, figure, blind text or accept list changed

• Lessons: L01 q3, q4 (stems from the list; same letters) · L02 q2, q3, q4, q5, q6 (from the list; q5's figure and blind text righted) · L03 q3, q4 (from the list; q4's figure and blind text righted), q6 (fb pair) · L04 q2, q3, q4, q5, q6, q7 (from the list; figures and blind texts righted; q7 options trimmed), q9 (new claim and scene).
• Attached: Summary Q2, Q3 (new scenes).
• Practice: PR-5 (from the list; pr_cliff_1 new colours, blind text changed), PR-10 (fb pair), PR-11 (from the list).
• Bank: PP-1, 2, 9, 14, 15, 16, 24, 26, 31, 35, 36, 41, 45 (was 46), 47 (was 48) built from the list (figures pp_cliff_1/2/3/4/5/6/9/10/12/14/15/8/16 — letter order shuffled on 1, 2 and 7; the inverted ones righted); PP-7, 8, 10, 12, 13, 19, 23, 25, 27, 28, 30, 32, 39 rewritten as above; PP-49 (was 50) option B; PP-45 (v1) cut, so v1's PP-46..51 are PP-45..50.
• Figures: every ITEM_CLIFFS key (28); cliff_five_headache, cliff_five_numbered, cliff_five_step1, cliff_fossils_headache, cliff_fossils_numbered (the one road cut); cliff_numbered, cliff_three_numbered (chips); no_sinking_card; landing_strip; karl_fossils_cartoon, julia_drawing_cartoon, roadcut_photo (briefs); eight keys deleted.

Sentence-atomise, the new sentences (one row each)

Judgement calls, v2

1. Lettering shuffled on about half the item cliffs, not all — the first drawn items of L01 and L02 keep alphabetical order top → bottom (a nine-year-old's first lettered cliffs read easiest that way); the bank's and L03 / L04's are shuffled so "bottom = last letter" earns nothing.
2. L01 grew by one sentence (351 words) for the scope line's contrast; the reviewer's "cutting the jar shortens L01" was a figure, so the article did not shrink. Kept whole, as recommended.
3. PP-32's new cliff puts a bone (tan sand) above a footprint (red-brown sand) with gray mud between — a dry-land story throughout, so the item tests the age read alone, not a reading of the place.
4. Topic 18's jar and the deposition card are no longer drawn here: L01 recalls both by sentence and item only (the figure review found both figures unplaced or misleading).

Owed

• A second non-author figure pass on the v2 renders is the coordinator's call (the v1 figure review's rows are all applied above).
• The blind read of blind/2026-10-02-v2/ (keys unverified until two cold readers agree; the thirteen righted items first).
• The picture pass (7 placeholders) after James's sign-off; James's word before the batch is sent.

v3 — polish (2 October 2026, after the v2 re-read)

Skill: course-build → write-mcqs (L04 q5, PR-2, L01 q3), write-explanation (the new fb and why lines), write-lesson (two L03 article lines), sentence-atomise (every new sentence) @ efd8fba

In plain words. The v2 re-read agreed every key (110 of 110) and left four soft flags. All four are applied. L04's story item has its own cliff, so its options no longer spell out the two items before it. The glacier in PR-2 is recalled in one line, in Topic 20's words. L03's article now says, at the burial strip, exactly what its quiz item and the bank's PP-30 rest on: the shell was buried when its own layer's pieces settled on it. L01 q3's cliff is cream, gray and red-brown, not the road cut's top three bands in the road cut's order. Every lettered item is still built from figures.ITEM_CLIFFS; the guard reads 34 of 34. Nothing is sent; nothing is committed.

Finding → change

Items whose stem, options, fb, why, figure or blind text changed

• L01 q3 (figure and blind text only; key unchanged) · L04 q5 (new cliff, stem, options, fb, why; key position unchanged) · PR-2 (stem and why; options and key unchanged) · L03 article (two idea lines at the burial strip).
• v2 → v3 key comparison on the sequestered keys: 109 unchanged, 1 changed (L04-05's key text, the new story). The type sheet's exemplars are untouched; TYPES_G4T21.md is unchanged.

Sentence-atomise, the new sentences

Checker results (v3)

• build_tools_check_topic.py → 0 FAIL, 8 WARN (the same three S13 lines and five S36 words, answered in v1).
• grade4/check_g4_articles.py → 0 FAIL (L01 353 and L04 349 WARN, answered in v1 and v2).
• check_svg_ids.py content/figures.py → PASS (56 figures, 0 ids, 0 shared).
• check_lettered_items.py → SUMMARY: 34 lettered items checked, 34 match their drawing, 0 mismatch.
• build_tools_build_topic_site.py → /tmp/g4t21_v3.html (4 lessons, 12 practice, 50 bank items, 56 figures). q_cliff_jkl and q_cliff_story_d rendered with headless Chrome and opened: the letters sit beside their bands; the footprint is in the top band, the roots in the soil, the shells at the bottom.
• Blind bundle blind/2026-10-02-v3/ (102 MCQ, 3 numeric, 1 FRQ, 4 order; 110 items). Its [Drawing: lines are colour and contents only.

Owed

• A cold re-read of L04-05 and PR-2 (the two items whose text changed) before the keys are called verified for v3; the other 108 stand on the v2 re-read.
• The picture pass (7 placeholders) after James's sign-off; the video batch on his review of the scripts.

Skill: course-build → review-lesson (Roles B, B2) → review-questions → review-accuracy (non-author pass; filed by the coordinator from the reviewer's hand-back) @ 3a2c39d

Topic 21 — guidance, questions and accuracy review, 2 October 2026

The teaching is right and the writing is clean, but the topic cannot ship as it stands, because ten item drawings letter their bands upside down. cliff() reads its letters list bottom → top; on ten figures the author wrote the letters top → bottom. The rendered drawing and the blind text agree with each other and disagree with the stems: four keys look wrong by the drawing (L02 q5 reorder; L04 q4; PP-32; PP-36) and nine more items name fossils in letters the drawing puts elsewhere (L03 q4; L04 q2, q3, q5, q6, q7; PP-15, PP-16, PP-35). The fix is one line per drawer — reverse those ten letters lists (pp_cliff_7 also its layer list) — and every stem and key then stands. The science is honest at Grade 4 depth; the quotes from Topic 18, Topic 20 and Grade 3 match word for word. Re-performed: scripts 150 / 149 / 150 / 150; intro 104; summary 150; articles 341 / 255 / 280 / 348 with a subtitle or figure every ≤3 idea lines; bank 51 → 13 / 13 / 12 / 13, keys A 9 B 13 C 13 D 11; blind lines colour and contents only; names clean.

The author's three questions, the reviewer's view: (1) stay term-free — yes (strata and superposition are grades 6–8's; cut PP-45, the one typed word item, which tests Topic 18's word); (2) keep "hundreds or even thousands of years" for one layer — yes (one honest order of size against "days"; fix PP-8's length tell); (3) keep L01 and L04 whole — yes (a figure or subtitle every three lines; cutting the jar shortens L01 at no cost).

Skill: course-build → review-figures (non-author pass; filed by the coordinator from the reviewer's hand-back) @ 3a2c39d — figures at g3-5-science commit 093047d

Topic 21 — figure review, 2 October 2026

The cliffs are clean, flat and in the stated order; shells, ferns, leaves, fish, footprint and bone read at a glance; the landing strip plainly shows each new layer landing on top. Then, on the L04 quiz, the stem says "Layer J holds fish fossils" and J in the picture holds leaves. Page pass: every figure renders on the built page as alone; check_svg_ids.py PASS. Verdict PASS once items 1–4 land.

1. Ten item cliffs are lettered upside down (cliff() reads letters bottom → top; these were written top → bottom): q_cliff_five_b, q_cliff_fossils_b, q_cliff_story_a/b/c, pp_cliff_4/5/7/9/10. Fix: one list feeds stem and figure (the brief's rule); add a guard that parses every "in layer X / Layer X holds" and bottom / top claim in stems, why and fb against the figure data.
2. The road cut is three different cliffs. L01 ROADCUT top → bottom red-brown, tan, gray, cream; L02 FIVE ("the same face") swaps gray and cream and adds a footprint; L03 FOSSILS puts gray with leaves on top while the text says "the gray stripe at the top" (L01 said red-brown). Fix: FIVE = ROADCUT + soil; FOSSILS = ROADCUT with the leaf in the top red-brown band; L03's line and karl_fossils_cartoon's brief say red-brown (one depiction per thing).
3. FORMED FIRST / FORMED LAST chips clip at the frame edge on cliff_numbered, cliff_five_numbered, cliff_three_numbered. Widen to ~430 or move the chips left.
4. no_sinking_card shows the state the misconception predicts (shell in the bottom band, nothing beneath). Tan floor, cream-with-shell, gray on top, so "stays in its layer" has a layer below it.

Unplaced (7): cliff_plain, landing_strip_headache, deposition_recall, story_table_headache, fern_fossil_photo, fish_fossil_photo, shells_slab_photo — delete all. cliff_headache and canyon_photo are video stills (keep).

Minor: pr_cliff_1 is pp_cliff_3 relettered (new colours); lettering runs alphabetically top → bottom on most fossil items ("bottom = last letter" is a learnable shortcut — shuffle a share); item frames ~33 % empty (tighten x0 when numbers are off); L02 "a thin dark stripe of soil" vs a full-thickness band; Topic 18's brown mud beside this topic's gray mud (a bridging caption word); adjacent same-colour bands on pp_cliff_8, q_cliff_story_c, pp_cliff_16 rely on a hairline; root glyph low-contrast; karl_fossils_cartoon brief colour and shell shape.

Three best: landing_strip, burial_strip, cliff_story_two with story_table_two. Three weakest: q_cliff_five_b, q_cliff_story_a, no_sinking_card.

Skill: course-build → review-lesson (Role A blind key read, two readers; Role C adjudication by the coordinator) @ 3a2c39d

Topic 21 — How rock layers tell the time — blind read, 2 October 2026

Keys did not survive. Both readers (111 items each) disagree with the key on the same set: one reorder key is stored in the wrong direction, and ten fossil items have stems whose layer letters contradict their drawings. Everything else agreed (Reader A 96 of 111 rows, Reader B 99 of 111; the three typed answers on the lists; the FRQ model in the key's words; the rule "lower is older; a fossil is as old as its layer" lands cleanly on every text-only item). The intro poll is not bundled by design.

The failed keys

Flags carried into the fix pass

Readers' files: blind/2026-10-02/READER_A.md, READER_B.md.

Correction after the guidance review (same day)

The guidance reviewer opened the rendered figures and found the root cause: cliff() reads its letters list bottom → top, and on ten drawers the author wrote the letters top → bottom (q_cliff_five_b, q_cliff_fossils_b, q_cliff_story_a/b/c, pp_cliff_4/5/7/9/10). The drawings and their blind text agree with each other and disagree with the stems; the stems and the sequestered keys stand (including L02-05's order, which is "first to last" against the drawing once its letters are righted). So the ruling above ("the stems' letters are the slip; re-key L02-05") is withdrawn: the fix is to reverse the ten letters lists (and pp_cliff_7's layer list), leave every key, and re-read the thirteen affected items cold against the corrected drawings. The vault rule still holds (one list feeds stem and figure), and the drawer's direction convention goes in the build record.

v2 re-read, 2 October 2026 (Reader C on blind/2026-10-02-v2, the whole bundle)

Every key agreed (110 items; MCQ disagreements 0; the four order items MATCH by check_order_keys.py — L02-05 M, K, P, L, N; PR-05 K, N, J, M; PP-24 N, J, P, K; PP-25 L, N, J; the three typed numbers on the lists; the FRQ model in the key's words). The lettered-item verdict: in all 34 items the stem's letter → contents pairs match the drawing and the letter sets are equal; the upside-down fault is gone. Soft flags to the polish pass: L04-05's option A spells out L04-02 and L04-03's answers on the same quiz page (move or re-cliff); PR-02's "glacier" is Topic 20's word (recall line); PP-30 option A is a strong tempter — confirm L03-02's sentence is on the lesson page; L01-03's colour run is three of the road cut's four bands (vary it); the bundle still prints an item numbered PP-45 after the cut (renumbered, fine). Reader's file: READER_C.md.