Week 3–10 deck: build note

DECK BUILD NOTE — Blocks A–E

Status: no block is cleared for recording. Five blocks passed QA with zero clean verdicts.

---

1. WHAT SHIPPED

Block	Weeks / days	Items	Sentences	Drills	Lesson days
A — response tokens, kosoado, first nouns	wk 3–4, card-days 1–14	168	120	14 (D1–D14)	14
B — wa/ga, particles, existence	wk 4–6, days 22–42	126	94	12	21
C — numbers, counters, time	wk 5–8, parallel to B and D	185 (header says 175)	84	8	20
D — verbs, -masu shell, i-stem, -nai stem	wk 6–8, days 36–56	130	102	8	21
E — the -te chant and its dependents	wk 9–10, 14 days	128	64	10	14
TOTAL	wk 1–10 (wk 1–2 no-card perception)	737	464	52	90 taught days across a 70-day calendar

Net new, honestly:

• A: 168 new. First card block, no inheritance. Intake is exact — 14 days × 12 items = 168, no day over or under. It is the only block engineered to the ramp.
• B: 126 carded, 52 are block-A duplicates → 74 net new. Six of those 52 B teaches before A does.
• C: 185 carded, no cross-block duplication (isolated domain). Item 101 (hyakugojuuen) is carded but scheduled on no day.
• D: 130 carded, 17 verbs duplicate B → 113 net new. iru/aru are declared re-cards (re-stored as pairs), which is legitimate.
• E: 128 carded, of which ~56 are D verbs re-carded as chant-row members — by design, not waste — plus 8 genuinely new lexemes (oshieru, dekakeru, wasureru, tetsudau, suwaru, aku, kekkon, furu) and the construction set.

Roughly 90 of the 737 slots are spent re-delivering material (excluding E’s deliberate chant re-cards). Net unique lexical inventory ≈ 612.

Sentence coverage is the weak axis. A: 48 of 168 items appear in no sentence card, ~30 of them marked PROD. B: ~30 of 126, including the entire hidari/migi/aida/mawari/ushiro/tokoro set and the PROD verb hanasu. D: 8 of 65 verbs (6 declared item-only, plus oyogu and shiru undeclared) — the best of the five. C and E card most items as drill members rather than sentence heads, which is defensible for counters and chant rows.

---

2. THE CORRECTIONS APPLIED

Course-wide, all five blocks:

• Modified Hepburn with doubled long vowels held everywhere. Zero -ei- spellings across 737 items. Zero macrons. Zero undropped/undoubled long vowels found by any of the five QA passes.
• Geminates doubled rigorously; kutsu, chichi, mitsu-type non-geminates correctly not doubled.
• n' before vowel/y correct in every environment that occurs: sen'en, ichiman'en, kin'yoobi, nan'yoobi, sanzen'en, ren'yookee.

Block A — yuu installed as citation (item 36, S29, S31, D4); the required i-adjective nai restored (item 168, capstone S120); all 92 nouns delivered in the X ga frame including the frame-only kango; me [1] and te [1] held 15 indices apart; collision senses withheld (hana-nose, ame-candy, kami-hair/god); soba, kawa, eki, kazoku made phrase-only; the asoko broken grid cell flagged out loud; the DO- column taught as one accent fact.

Block B — yuu (77), au [1] (76), nai as both i-adjective and aru’s suppletive (36/38), soba/kawa permanently phrase-only, kami-hair excluded, te/me 17 indices apart, all five kango carded only in collocation, the family row taught as an in-group/out-group paradigm rather than ten words.

Block C — every n' environment correct; too not *tokoo; yonin as three morae; ichiman not *juusen; yoji/kuji/shichiji/shigatsu marked lexicalized; ree not rei; the yokka/yooka/tooka length triad isolated as its own drill.

Block D — the single densest correction set. yuu/yutte/yutta plus iwanai not *yuwanai (items 5, 104, S72); au [1] with atte homophony made permanently phrase-only (47); nai [1] present in both roles (107, S76); kookoosee and isshuukan audited at 6 morae; -masen accent corrected to SE, -mashoo to SHO (84, 87) with any *nomímasen restarting the drill run; w-insertion restated for all four vowel shapes (-au/-iu/-ou/-uu) because the textbook -au statement mis-generates for yuu (103); shinu named as the language’s only -nu verb; the -eru/-iru trap neutralized by storing class rather than computing it.

Block E — the fatal earlier omission fixed: -chau and -jau are carded in the same session, both allomorphs, error-gated to zero voiceless leakage (102–107, D9). yuu citation + tte yutteta (71, 128). au [1] applied, and the earlier pass’s false claim that pitch separates atte(au) from atte(aru) explicitly withdrawn — separation moved to argument frame (51, 65). R0 (ichidan) and R6 (irregulars) restored, both omitted entirely by the earlier pass; ichidan is a third of the ledger. -ta shipped in the same session as every -te row rather than deferred. shiranai flagged as the one place -tenai does not apply, carded in the same breath as the rule it breaks. -naide kudasai and -nakute mo ii carded beside their -te twins to block *-te nai kudasai.

---

3. WHAT THE CONTINUITY AUDIT BROKE

The arithmetic does not close.

• Week 4 busts by 31% — 110 cards against an 84-card ceiling, because A and B run concurrently for seven days.
• Week 6 closes only because 17 of its cards are redundant. Delete the duplicates and it fails.
• Eleven individual days exceed the ceiling. Meanwhile week 7 has 40 spare slots and block E has 124. The intake is spiked where blocks collide and starved where one runs alone. This is not a rounding problem; it is what parallel construction did to the ramp.

29 items are assumed by at least one block and delivered by none. The parallel design produced exactly the predicted failure — each block assumed a neighbour owned the word. The worst: ii (needed by B item 112’s own chunk tenki ga ii, by D, and by E’s yokatta), watashi (B’s model double-nominative card S62), gakkoo, sensee, biiru, koohii, kippu, shatsu, zasshi, itsu, demo, plain past -ta (demanded by a B drill on day 25, delivered week 9), and the i-adjective inflections -kute/-katta (used by E’s S14 and S56, taught nowhere in the course).

Duplication. Block B spends 41% of its budget re-teaching block A — 52 of 126 items — including six it teaches first. Three re-cards disagree with A on the data:

Item	Block A	Block B	Correct
shigoto	[0]	[2]	[0] (B is wrong)
kami ‘hair’	[2], withheld	[1], “deliberately excluded”	[2] (B’s note is wrong twice)
eki	kango, phrase-only forever	bare eki ga	A’s constraint, broken

D then re-teaches 17 of B’s verbs.

Five inherited constraints are broken by the block that inherited them. eki (A: phrase-only → B cards it bare). nihongo (A: frame-only inside #38 → D uses it bare four times). kawa (B declares it permanently phrase-only at item 59 → B’s own S43 is Kawa ga aru). soba (A’s frame X no soba → B invents kawa no soba ni, building the permanent frame on the other permanently-framed collision item). jikan (B: collocation-only → D’s S1 breaks it).

The contraction promise is passed forward three blocks and kept by none. A’s day-14 note promises B will deliver -teru, -chau/-jau, -toku/-doku, -tara/-dara, -cha dame/-ja dame with both allomorphs. B delivers zero contractions. D defers all of them to E. E delivers exactly one family — -chau/-jau. The remaining three pairs are re-deferred to weeks 11–12, which do not exist as material.

Interference separation is enforced by three incompatible mechanisms and violated in four blocks. A’s 15-index rule is honoured in A (me/te) and broken everywhere else: B cards hidari/migi at 55/56, mono/koto at 100/101, uchi/ie at 23/24; C cards two items with identical romaji, identical accent and identical audioText at 94 and 97 (rokkai floor / rokkai times) — an unanswerable card in an audio-only deck; E cards yonde [1] and yonde [0] three indices apart inside the audio that induces the row, and cards atte/atte fourteen apart while its own note says never to drill them adjacently — then places S13 and S14 adjacent in the same session.

The induction base closes on the last possible day, and block E then breaks it.

Closure by teaching day (not by frequency index — D’s own notes confuse the two, and its day-42 audit claim “-ru class already closed at toru” is false by fourteen days):

Coda	Closes	On
ichidan	d42	neru
godan -ru	d43	hairu/hashiru
godan -u	d46	morau
godan -ku	d46	hataraku (regular members only)
godan -su	d49	naosu
godan -gu	d49	sawagu
godan -bu	d50	erabu
godan -mu	d56	tanomu
godan -tsu	d56	tatsu
godan -nu	never	shinu — 1 member in the language, correctly memorised

-mu holds two members and -tsu holds one until day 56, D’s audit day, which introduces 13 verbs. Both classes reach threshold with zero consolidation.

Then E’s CHANT-7 call sheet drops katsu, utsu and suu — three verbs admitted to D specifically to fill coda classes. Counted inside the drill that actually induces the rows, the -tsu coda has three members (matsu, motsu, tatsu), below the stated threshold. E also drills tanomu with no E item card, and its ASPECT SORT requires six -teru forms (tsuiteru, haitteru, yonderu, nonderu, asonderu, hataraiteru) that exist nowhere as cards. The handoff is described with three different numbers across two blocks: D says 65 pairs, E’s header says 60, E’s chant uses 61.

The wa withholding: content passes, schedule breaks it. All 120 A sentences are clean — no topic wa. The two frozen forms are correctly flagged (konnichiwa #25, konbanwa #26). One unflagged edge case: S40 soo de wa nai ships the etymological topic particle in a card labelled only “the formal expansion of ja nai” — RECOG-only and never analysed, but A’s own flagging convention was not applied to it. The real failure is the calendar: B introduces wa on day 23; A’s last sentence card is day 28. For six days A scrupulously avoids a particle B is actively drilling. A’s item 66 rationale — “topic wa is deliberately withheld to block B” — is void for the last 43% of the block.

Two things close cleanly. A’s intake arithmetic is exact. A’s wa content withholding holds across all 120 sentences.

---

4. WHAT IS STILL WRONG

4A. BAKES INTO AUDIO — fix before any TTS batch

These are accent integers and audioText strings. Once recorded they are expensive to unpick.

1. Block D, the entire plain-negative accent family — ~35 integers, all one mora late. D states the correct rule (“the accent lands on the mora immediately before -nai”) and then indexes onto the na of -nai. nomanai is [2] not [3]; wakaranai [3] not [4]; tabenai [2] not [3]; minai [1] not [2]; shinai/konai [1] not [2]. D’s own hedge on konai (“some references list [1]”) is the tell. This is the most-drilled form family in the block. Severity: fatal.
2. Block E, six composite -ta cards, 15 wrong integers. Items 27, 28, 29, 36, 48, 77 stamp one integer on sets whose members carry two or three different accents. Item 77 gives [0] to katta/matta/totta/itta/shita/kita — matta, totta and kita are [1], which destroys the kita/kite contrast item 75 calls load-bearing. Fatal.
3. Block C, items 94 and 97 — identical stimulus, two required answers. Unanswerable in audio. Fatal.
4. Block E, nete [0] → [1] (neru is [1]); propagates into the R0 chant row and S5/S9.
5. Block E, -chau/-jau accent. Fixed [3] is right only for four-mora stems. The rule is accent on the cha/ja mora: wasurechatta is [4], hanashichatta [5]. And the block’s accent law needs its one exception stated — -chau/-jau is the only shell that moves the host accent (nomu [1] → nonjau [3]).
6. Block E, shiranai [3] → [0]. Heiban verb, heiban negative.
7. Block A, hako [1] → [0]. Takes drill D9’s minimal pair down with it — as carded, the learner is trained to hear a distinction not present in the audio.
8. Block B, shigoto [2] → [0]; kami ‘hair’ [1] → [2] (and it is therefore accent-identical to kami ‘paper’, which B’s note gets backwards).
9. Block C, the numeral accent table is internally incoherent. juuichi [4] → [3]; sen'en [3] → [1] (the deck’s most-heard money word); nihyaku [1] is likely [2]; roppyaku/happyaku/hassen shift accent while rokujuu/hachijuu do not, with no rule stated. fukuro [0] → [3]. Re-derive the whole 20–9000 table against one authority.
10. Block A, kane → okane at the block capstone. kane ga nai drilled cold installs a register error at the highest-salience point in the block.
11. Accents that cannot be tested by ear as recorded. A’s ohayoo [4], konnichiwa [5], konbanwa [5] — odaka on forms that never take a following particle. C’s odaka numerals and -tsu counters recorded bare (ichi [2], mittsu [3], too [1], futari [3], roppyaku [4]). B’s kesa/ima bare. D’s mainichi bare. E’s kekkon outside the ga-frame. Either supply a carrier or drop the integer.
12. PROD cards shipped with no accent. D items 92, 93, 97, 98, 99 (-ni iku, -kata, -hajimeru, -owaru, -tsuzukeru) carry - plus “verify before recording”. E items 34, 90, 91, 92 say “accent approximate”. A PROD card with no accent is not recordable.
13. Multi-word chunks with a single meaningless integer. A #9, #14, #28, #37, #38, #39, #40. moo ichido onegai [1] is two accent phrases, neither of which is [1].
14. Mora counts misstated in notes that drive timing. A: chikaku “4 morae” (3). C: kokonotsu “5” (4), tsuitachi “5” (4), reshiito “5” (4). D: suwanai “su-u-wa-na-i” (4). E: kekkon shiteru “5-mora collocation” (7).
15. Kanji in the deck — contract breach, two occurrences. Block D item 5’s note and day 41’s drill text both print 言う. Delete before the text goes near a TTS pipeline.
16. Block A, honto carded as base under a doubling contract; hontoo is the base and honto the labelled clip.

4B. BLOCKING BUT NOT AUDIO — fix before a learner sees it

• 1T is unverifiable across the whole course. No block declares its inherited inventory. C runs parallel to B and D and declares nothing, so all 84 of its sentence cards fail mechanically. E declares only function words and runs ~30 uncarded content words. Nothing can be certified until one cumulative inventory exists.
• Unlearnable cards. B item 112 (tenki ga ii — ii exists nowhere). B item 38 (teaches nai on d30 using jikan from d40). C item 101 (carded, never scheduled). D naru (carded, drilled, never taught on any day). A kawa (declared frame-only, never receives its frame in any sentence). A jaa (used at S20, carded nowhere in 168 items).
• Drills scheduled against inventory that does not exist. B: six of twelve unrunnable as placed, two reaching 8–18 days forward. C: six drills demand ~12 uncarded items. D: CODA-CLASS SORT opens d42 with 13 of 14 items unintroduced. E: day 5’s permanent full randomisation breaks the R4/R5 separation rule, the two-blocked-sessions rule and the IKU ISOLATION rule simultaneously; the NASAL-VS-SILENCE gate that decides whether R5 opens on day 4 is scheduled inside the day-4 session that opens R5.
• The -de allomorph is systematically absent from eight E constructions. -te kudasai, bare -te, -te kureru, -te morau, -te kuru, -te kite, -te mo ii, -te ii are shown only on voiceless hosts. No yonde kudasai, no nonde mo ii, no isoide kite. This is the same one-sided-pair defect E congratulates itself on fixing for -chau, reproduced eight times.
• Two core A cards are not Japanese: bare kore ga as an answer to dore?, and koo koo where a speaker must say soo soo. B S22 Kono ie wa mado ga aru is not said. D S98 okane o motsu requires motteru. E S7 kutsu o nuide, ie ni ita is incoherent.
• Structural holes. A teaches 90 nouns, 10 relational nouns and ga with no aru, no iru, no ii, no wakannai — one verb and one adjective in 168 items, so nothing in its second week can be predicated. B ships ue/naka nowhere while using shita uncarded. E omits -te kara, -te miru, -te ageru, -te iku, -te aru and the entire i-adjective inflection it uses.
• A false rule that will train a wrong form. C items 149/150 declare dates regular from 21 on; 24th is nijuuyokka, 29th nijuukunichi, 17th juushichinichi, 19th juukunichi. As written the deck trains *nijuuyonnichi.

---

5. THE BUILD ORDER FOR THE TOOLING

Nothing below step 0 can be validated until step 0 exists.

0 — One inventory, one calendar. A single cumulative item table keyed on canonical romaji, plus one shared calendar mapping every block-day to an absolute course-day. This is the prerequisite for the 1T checker, the ramp checker, the collision index and the duplicate merge. It does not exist today and it is the reason four of the five continuity failures were invisible to the block authors.

1 — Merge and reconcile. Delete B’s 52 A-duplicates and D’s 17 B-duplicates (this alone resolves the week-4 and week-6 arithmetic). Resolve the three data conflicts (shigoto, kami, eki) to a single record. Mark E’s ~56 chant re-cards as recard: true so the ramp checker does not count them as new intake.

2 — Lock the accent table. One authority, one edition, cited per item. Re-derive from scratch: D’s ~35 -nai integers, E’s six composite -ta cards, C’s full 20–9000 numeral table, and every item currently carrying -, “approximate” or “verify”. No item ships without an integer or an explicit accent: null, reason: unaccented-clitic.

3 — Generate audioText from the frame rule, not by hand. One function, one table:

Class	Frame	Example
noun	X ga	neko ga
relational noun	N no X ni	isu no shita ni (not shita ga)
kango	mandatory collocation	jikan ga nai, eki no mae
verb	dict, te pair	nomu, nonde
numeral / counter	X ga or a carrier	mittsu ga, sanji desu
particle / clitic	phrase	kore ga, sanji kara
chunk	whole, per-phrase contour	moo ichido onegai
interjection with no particle slot	carrier phrase or accent: null	ohayoo

Fields per item: id, block, day, romaji, accent, pos, gloss, cardType, audioText, teForm, frame, formalExpansion, allomorphPartner, collisionSet, prereqs[].

4 — Validate before recording. Every check is mechanical; all must pass.

1. Every sentence’s newElement count is exactly 1, and every other token resolves in the cumulative inventory at an earlier absolute day. Same-session items do not count as prior.
2. Every drill’s stimulus list resolves at or before the drill’s scheduled day.
3. Every item is introduced on exactly one day, and every scheduled day’s introduces list is non-empty and consistent with its sentenceRange.
4. Every item declared PROD has at least one sentence card, or is explicitly demoted to RECOG.
5. Every declared phrase-only item appears only inside its declared frame, in items, sentences and drills alike.
6. Every accent-identical or segmentally-identical pair is ≥15 indices apart and on different days and not co-scheduled by any drill.
7. Every allomorph pair has both members in the same session: -chau/-jau, -toku/-doku, -tara/-dara, -cha/-ja dame, -ite/-ide, -te/-de hosts for every dependent construction.
8. Every long vowel doubled, every geminate doubled, n' before vowel/y, no kana, no kanji, no macron — regex, run over every string field including notes.
9. Every stated mora count matches a computed mora count from the romaji.
10. Daily and weekly card totals against the ramp ceiling, counting re-cards separately.

5 — Record. Human voice, not TTS, for:

• Every collision set and accent minimal pair. kite/kite/kiite/kitte(verb)/kitte(noun); yonde/yonde; katte/katte; atte/atte; the three heiban itte; shite/shitte; kaete/kaette; ite/itte; sangai/sankai; nihon/Nihon; yokka/yooka/tooka; sanbon/sanbiki/sanbai/sanpun. These are the deck’s product; a synthesised near-miss is worse than no card.
• Everything the QA flagged for verification — D’s five compound i-stem tails, E’s four approximates, konai, tabemono/nomimono, kite [1], sen'en, fukuro.
• All prosody-carrying cards. A’s rising terminal (#68), ne/nee, daijoobu rising vs falling, soo desu ka falling vs rising, trailing chotto; E’s manner--te (no pause) vs sequence--te (pause) — the pause is the contrast; D8 in full.
• All perception-drill stimuli. E’s NASAL-VS-SILENCE (and fix asonde/atatte → asonde/asotte first — as written the pair is not minimal), A’s D1 un/uun, A’s D9 X-ga discrimination, C’s length triad.
• Every multi-word chunk with two accent phrases, and every contracted form: -ten no?, yutteta, -chatta/-jatta, the particle-drop careful/stripped pairs.
• All conversational-speed material: A’s D14 four-minute shadow, C’s transaction shadow (re-sequenced — as written the price is quoted twice and the learner decides after being quoted), E’s S63/S64, D’s day-56 audit run.
• Every metronome call sheet — E’s CHANT-7 at 60/72/84/96/108 bpm needs one voice at consistent tempo, with a human check pass on the 500 ms confirmation tokens.

TTS is acceptable only for single-word items with a locked, verified integer, no collision partner, in the correct frame, at citation tempo — and only after step 2 closes. Batch by accent class, spot-check every batch against the human reference for that class.

Recording gate: no batch goes out until checks 1–10 pass and the accent table is signed off. The current state fails checks 1, 2, 4, 5, 6, 7, 8 and 9.

---

6. WHAT COMES NEXT

Weeks 11+ already owe, from promises made in A, B, D and E:

• The three remaining contraction families with both allomorphs: -toku/-doku, -tara/-dara, -cha dame/-ja dame. Plus -nakya, -tte, -n da/nan da/soo na n da (E cards only the rising-no question side).
• The -ta branch E’s own defence sells as the chant’s second dividend and then never ships: -tari, -tatte, -ta koto ga aru, -ta hoo ga ii, -ta ato de, -ta bakari.
• The remaining -te dependents: -te kara, -te miru, -te ageru, -te iku, -te aru.
• The i-adjective class in full. It has no home block. ii, -kute, -katta, -ku naru, yokatta — all four are used by existing cards and taught nowhere. This is the single largest hole.
• naru, transitivity pairs, potential, plain volitional, conditionals (-ba, nara, to), shitteru/shiranai as a paradigm, wakannai, iku/kuru deixis, a keigo floor.

Decisions required before a line of week 11 is written:

1. Serial or parallel. Parallel construction caused every continuity failure in this audit — the undelivered 29, the 52 duplicates, the week-4 bust, the wa schedule break, the split induction base. Decide, and if parallel survives, the shared inventory in step 0 becomes mandatory infrastructure rather than a repair.
2. Who owns the cumulative inventory, and is it the single source of truth the 1T checker runs against? Without a named owner it will drift again by week 12.
3. Is the 84/week ceiling real, and does it count re-cards? Week 6 currently passes only by counting redundant cards toward the quota.
4. Register default. D says produce -masu, comprehend both. E teaches bare -te as the unmarked request and -te kudasai as marked. These contradict. Pick one and rewrite the loser.
5. Accent authority — one dictionary, one edition, cited per item, and a rule for what to do when it disagrees with the target casual register.
6. Collision policy. Three mechanisms are in use (15-index separation, day separation, permanent phrase-only) and none is applied consistently. Choose one primary with the others as escalations.
7. Where the i-adjective class lands, and whether it can be retrofitted into weeks 5–8 or must open week 11 — because -kute/-katta are load-bearing for E cards already written.
8. Whether recognition-only production deferrals are tracked. E defers kiru(wear), kiru(cut), katsu, kaeru(change) to weeks 12–14. Nothing currently holds that list.
9. Does the week 1–2 perception block still exist as runnable material? E’s exit gate instructs a return to it on gemination failure. If it is not maintained, that gate is decorative.
10. Voice count and mixing policy — whether human and TTS tokens may appear inside one drill run, and whether a single voice covers the whole course.

Fix order for what already exists, if this is repaired rather than rebuilt: (i) publish the shared inventory and calendar; (ii) delete the 69 duplicates; (iii) card the 29 undelivered items, ii and watashi first; (iv) lock the accent table, D’s -nai family and E’s composite -ta cards first; (v) return katsu, utsu, suu to E’s R5 and move motsu, tatsu, yasumu, sumu out of D’s day 56 into days 44–48; (vi) re-space the -te homophone family and break up E’s day 4.