Completeness critique

1. Ordering questions a serious learner still has that this does not answer

Within-stage order is completely absent, and that is where most of the ordering lives. Stage 17 is a six-week block containing -soo, mitai, rashii, -ppoi, soo da, deshoo, kamo, hazu, beki, -garu, direct passive, adversative passive, possessor passive, honorific -rareru, ~30 transitivity pairs, and -te aru. That is roughly forty items presented as a set. The document’s entire thesis is that sets are not orders. Same problem in stages 12, 13, 16 and 18. The stage boundaries are ordered; the contents are not.

The order of the 35 -te dependents. The whole justification for the -te budget is fanout. Having spent that argument, the document ranks exactly one child (-teru first) and then scatters the rest across stages 12, 13, 16, and 17 with no application of its own fanout×frequency rule. Where does -te mo ii sit relative to -te miru, -te oku, -te kara, -te ageru? Unanswered.

The order of the first 300 vocabulary cards. “Randomized by conversational frequency rank” is an ordering abdication dressed as a principle. The document itself orders the twelve seed chunks by phonological difficulty in Stage 1 — and then abandons that principle for the remaining 5,988 items. Should 2-mora words precede 4-mora ones while the beat grid is three weeks old? Should heiban nouns (roughly half the noun lexicon) lead, or should accented ones lead so the drop is established as a category before the flat pattern is trained as the default? These are exactly the item-level questions asked, and none is addressed.

Which accent class first, and does accent class order interact with mora count? Accent is tagged from card 1 but never used as an ordering variable, only as a card-format constraint.

The order of the five rows inside the -te chant. By the document’s own fricative-before-stop logic (train the audible cue before the silent one), -mu/-bu/-nu → -nde should precede -u/-tsu/-ru → -tte, because nasal voicing is audible and gemination is a beat of silence. The document has the argument and never applies it to the chant it calls the largest single budget in the course.

Which Sino-Japanese roots, in what order. N+suru is gated on “40 Sino-Japanese nouns” and the multiplier is claimed over “800-1000” of them. Not one is named, and no ordering rule for them is given beyond “no two competing readings within 200 items” — which at N=20 is ten days, flatly inconsistent with the five-week separation demanded for -eru. Which is the rule?

Failure branches. Every stage has a pass criterion; only Stage 2 has a failure branch. If the 300 ms -te criterion is missed at week 11, what is the reorder? What is the resume order after a three-week lapse at month 5? A critical path with twenty gates and no slack model is not a schedule.

Lexical prerequisites per grammar stage. Only Stage 6 states one. Transitivity pairs need ~60 specific verbs already stored; evidentials need an adjective inventory; the benefactive triangle needs the giving/receiving argument frames. None of these dependencies is stated, though the document’s central mechanism is dependency.

2. Where it is still operating at the level of categories

• “~180 concrete high-frequency nouns” (Stage 3). The single largest early card block. Zero items named.
• “First 60 verbs by conversational frequency” (Stage 6). Not listed. This one list is the input to the -te gate, the -nai gate, the class-assignment drill, the potential stage, and the transitivity stage. It is the most load-bearing list in the document and it does not exist.
• “~60 core adjectives, randomized” (Stage 10). Not listed. The “about three” exceptions to the vowel+i rule are not fully enumerated (the real list by ear includes kirai, yakkai, yukai, saiwai, at minimum).
• “~35 direct dependents” of -te. Never enumerated. This is the number that wins the largest scheduling contest in the course, and it is uncheckable — see §4.
• The 140-item exit battery. Grammar terminates on this test, so the battery is the operative definition of the closed set. It is given as sixteen category names.
• The counter block. Ten counters named; three chanted. The -hiki and -hai series (ippiki/sanbiki/roppiki/happiki, ippai/sanbai/roppai/happai) are exactly the same p/b/h alternation and are never given. Yonin — the actual irregular in the -nin series — is missing while hitori/futari/sannin/nannin are listed. Sanzen is missing from the thousands. Ordinals (-me, -banme: “futatsume no shingoo”) are absent entirely and are unavoidable in directions.
• Kosoado “3 exceptions flagged” — none of the three is named, and the doko cell is dropped from the row listing.
• “Question inventory INCLUDING nande and dooshite” — the inventory itself is not given.
• “~60 memorized compound tails” — twelve named.
• The i-stem tail set. Law 4 claims naming the i-stem converts twenty later grammar points into one operation plus twenty tails, and lists them: -tai, -nagara, -ni iku, -kata, -sugiru, -yasui, -nikui, -hajimeru, -dasu, -owaru, -tsuzukeru, -naosu, -komu, -soo, -nasai, o-V-ni naru. No stage ever schedules any of them except -tai (Stage 11) and -soo (Stage 17). -nagara, -kata, -ni iku, -sugiru, -yasui, -nikui, -hajimeru, -owaru, -tsuzukeru simply never appear in the sequence. Likewise the ka/mo/demo indefinite paradigm (dareka, nanika, dokoka, itsuka, nandemo, daredemo) is invoked twice as a “multiplier” and never scheduled. Likewise purpose clauses — tame ni, yoo ni, yoo ni naru, yoo ni suru — which are core and appear nowhere. Likewise mama. A curriculum whose claim is completeness of ordering has roughly twenty-five named items with no position.

3. Remaining factual errors

Mora arithmetic. kookoosee is six morae (ko·o·ko·o·se·e), not the (5) given — and the error sits inside the document’s own beat-notation contract. -teta / -tetta: the counts (3) and (4) match neither the suffix (te·ta = 2, te·Q·ta = 3) nor the cited words (futteta = 4, futtetta = 5). byooki (3) vs biyoki (3) is placed in a length tier but is not a length contrast at all — both are three morae; the palatal contrast is carried correctly later by byooin (4) / biyooin (5), making the earlier pair either redundant or wrong. Ten counters × seven cells is 70, not “~56”. anai in the moraic-N set is not a word (fine for HVPT, but it is presented as a controlled lexical pair).

“-masu forms are uniformly accented on ‘ma’ without exception.” False, and the “without exception” makes it worse. -masu and -mashita take the accent on ma; -masen takes it on se (nomimasén); -mashoo takes it on sho (ikimashóo). Three of the five shell endings the document teaches at Stage 6 violate the rule it states there.

“Tookyoo daigaku” as compound deaccenting. Nihongo [0] is genuine deaccenting. Tookyoo daigaku is not — a three-plus-mora accented second element carries the compound accent onto its own first mora. The example contradicts the phenomenon it illustrates.

The -eru disambiguation is broken on its own examples. Law 5 and the contested call both state: “potential takes ga, transitive takes o.” That works for kaeru (can buy / 変える). It does not work for kireru, yakeru, toreru, wareru, nukeru — every one of which is a potential colliding with an intransitive, not a transitive. Rōpu ga kireru (the rope snaps) and rōpu ga kireru (I can cut the rope) both take ga. The claimed surviving cue does not survive on five of the six cited pairs. And worse: the document argues at Stage 3, Stage 4 and Stage 18 that particles are dropped at high rates in casual speech and trains the learner to expect the elliptical form as primary. The one disambiguator it rests the -eru schedule on is the one it has trained the learner not to expect.

Transitivity taxonomy misfiles its own key class. The batching lumps kieru/kesu, deru/dasu, and hieru/hiyasu into one “-eru/-asu” class; they are three different alternations (kes-u, das-u with root vowel change, hiyas-u). More seriously, the taxonomy omits the intransitive -eru / transitive -u class entirely — yakeru/yaku, kireru/kiru, wareru/waru, nukeru/nuku, oreru/oru, ureru/uru — which is precisely the class that generates the potential collision the document spends a contested call on. The class is discussed in one section and absent from the taxonomy in another.

“shitteru’s negative is shiranai, suppletive.” The observation is right; the label is wrong. Shiranai is not suppletion (compare aru/nai, ii/yoku, correctly labelled elsewhere) — it is a lexicalized preference for plain negation over resultative negation. In a document that corrects other people’s terminology (rendaku, regressive assimilation, “-ta kara”), this is a self-inflicted wound.

The obligation ladder is not a ladder. “-tara doo < -ta hoo ga ii < -nakya < -nakereba naranai < beki” places -nakya below -nakereba naranai, but -nakya is -nakereba contracted. They differ in register, not deontic strength. The document’s own Law 3 says so.

Speech-rate claims are probably inverted. Conversation is put at 420-480 morae/min and “news/announcer pace” at ~350. Read news is generally at least as fast as spontaneous conversation; spontaneous speech carries far more pause and filler time. More importantly, gross morae/min conflates pause with articulation rate, so the entire ladder (300 → 330 → 350 → 380 → 420 → 450) is denominated in a unit that does not measure the thing that breaks the parse. Local articulation rate or syllable rate excluding pauses would.

Missing, and glaring for an audio-first plan: iu is pronounced /juː/. 言う is yuu; 言った is yutta; って言ってた is tte yutteta. The document leans on -tte harder than on any other single item, correctly notes iu/iwanai in the w-insertion rule, flags itte as a two-owner homophone — and never mentions that the citation pronunciation of the source verb is not what the learner will hear. This is the exact class of error the document was built to prevent.

Its own romanization contract is violated throughout. The contract says long vowels are always doubled and Stage 10 explicitly notes that kirei, teinei, yuumei are heard as /-ee/. The document then writes kirei, teinei, yuumei, gakusei, sensei, eiga, keizai, and chou. If romaji is the only notation the learner has, this is not cosmetic.

Unflagged interference the document’s own Law 5 should have caught: ageru 上げる (transitivity pair, Stage 17) vs ageru あげる (benefactive, Stage 13) — same accent, four months apart, no ledger entry. Same for kureru 暮れる / くれる.

4. Unfalsifiable or circular ordering claims

The fanout metric is circular. “-te has ~35 direct dependents” is a count over a graph the author drew, using a notion of “direct dependent” defined by the curriculum. The document even performs the circularity in public — it revises 45 down to 35 because the higher figure “double-counts second-order nodes hanging off -te iru and items taught inside the -te unit itself,” which is a statement about how the author chose to unitize, not about Japanese. Then Stage 8 reverts to “45 constructions are waiting.” Neither number is checkable, and the largest scheduling decision in the course rests on their ratio to “the runner-up,” which is also never named.

“Perception gates encoding” does not gate. The exit is tiered such that a learner below 75% at four weeks “releases anyway.” A gate that admits everyone is a schedule, not a gate — and the accompanying mechanism (“an SRS review strengthens the trace you had at encoding rather than repairing it”) is stated so that no outcome disconfirms it: success means the block worked, failure means encoding was corrupted, a learner who skips it and succeeds “already had the categories.”

The review-load law contradicts its own sanity check. The formula given is reviews/day = N × (number of ladder steps). Applied to the stated example — ladder {1 day, 100 days}, N = 1 — it yields 2 reviews/day. The document says “reality after 1,000 days is about 10.” Ten is what you get if the terminal step repeats forever, in which case the quantity grows without bound and there is no steady state to have a constant for. The corrected arithmetic is probably right in direction (the constant is the number of steps reached, which grows logarithmically with program length), but “6-7N” is a fitted number over an unstated horizon at an unstated retention rate with lapses ignored, presented as a law. With realistic lapse rates the multiplier is nearer 8-10N, which breaks N=20 too.

The vocabulary ledger does not survive its own cap. Cards open week 3 at 10/day. By end of week 4 that is 140 items; Stage 3 claims ~280. By end of week 7 the maximum under the stated ramp (14×10 + 21×15) is 455; Stage 6 claims 800 in its vocab line while its trigger demands ≥500. The grammar gate — the single most emphasized gate in the document — is unreachable at its stated date under the document’s own N cap. Separately, Stage 5 runs “to ~700” in weeks 5-8 in parallel with Stage 6’s “to ~800” in weeks 6-7, which either double-counts or conflicts.

“Five weeks” of separation for the -eru collision is asserted with no mechanism and is inconsistent with the “no two competing readings within a 200-item window” rule used for kango, which at N=20 is ten days.

The tuning-rule trigger is nearly unreachable: “if comprehension does not improve for six weeks while vocabulary grows by 1,000+” requires ~24 cards/day against a 20/day cap.

The gates are statistically underpowered. ≥18/20 and ≥16/20 thresholds on 20-item batteries carry a binomial standard error near 10 points. Half the exit tests cannot distinguish pass from fail.

Every “ifWrong” is asymmetric by construction — mild cost if this plan errs, catastrophe if the alternative had been taken, ten times out of ten. That is a rhetorical signature, not an analysis. And the epistemic standard is applied unevenly: a source’s -toku frequency claim is rejected as “unsupported,” while “-chau is clearly more frequent than the passive” is asserted with nothing behind it and then used to cut passive instruction.

5. The strongest argument against this ordering, steelmanned

A competent pedagogue would not attack any individual placement. They would attack the premise that placement is the high-variance term.

The argument: instructional sequence is a small-effect variable and input volume is a large-effect variable, and this plan systematically trades the large for the small. Thirty years of morpheme-order studies and Pienemann’s teachability hypothesis converge on the finding that acquisition order is substantially internally determined — learners acquire the resultative reading of -te iru late regardless of when it is taught, and instruction advances a structure reliably only when the learner’s processor is already at the immediately preceding stage. If that is right, then the marginal return on getting -tara six weeks before -eba rather than adjacent is small and possibly zero, while the plan’s active-input allocation — 12 minutes/day at week 3, 20 at week 11, 32 at month 7, uncapped only at month 9 — accumulates something on the order of 100 hours of attended listening in the first nine months. A cruder plan spending 15 minutes on cards and 75 on graded audio from week 3 would accumulate three to four times that, and would very likely have better listening at month 9 despite a worse order.

The plan’s defense is Law 9: below ~95% lemma coverage input is “decoration.” But that threshold is imported from reading research (Hu & Nation, Laufer), and its transfer to listening is contested — listening carries prosodic, redundancy and interactional cues that reading does not, and reported thresholds for spoken comprehension are lower and softer. The plan asserts the reading threshold as a hard gate and uses it to justify capping the one variable with the best evidence behind it.

Does it land? Partially, and it is the objection the document is least equipped to answer.

It lands on the input allocation. The 95%-coverage cap is the weakest-supported load-bearing claim in the document, and the schedule it produces — eight months of listening as a side dish — is the plan’s largest exposed risk.

It lands on the productive claims. Nothing about plain-form production at month 9 or register control at month 24 follows from a dependency graph; those are precisely the outcomes processability predicts are sequence-insensitive.

It does not land on the interference work. A-B/A-C retroactive interference is among the most robust findings in verbal learning, and the homophone-spacing, contrast-pairing, and thematic-clustering decisions are the best-supported parts of the document. Nor does it land on the reduction-as-citation-form call, which is a claim about what the input contains, not about acquisition order, and is straightforwardly correct.

Net: the ordering is defensible; “order of acquisition is everything” is not, and this document would be stronger if it conceded that the order buys perhaps 20% and the input volume buys the rest — and then re-derived the daily braid under that concession.

6. The single addition that would most improve it

Emit the actual ordered ledger. One flat, numbered list — one row per item, in sequence — with four columns: index, item, the gate it is behind, and its interference partner with the minimum index separation. Not a policy for generating an order. The order.

Minimum viable version, if only one piece can be produced: the ordered list of the first sixty verbs, each tagged with class and coda, showing which four-per-coda-class satisfy the Stage 6 induction base and at what index each is reached. That single list is the input to five separate gates in this document and is the one thing a learner cannot reconstruct from what is written.

The reason this is the right addition rather than any content fix is that producing it is also the falsification test. Writing the ledger out forces the week-7 500/800-word gate to collide with the N-ramp; forces the twenty i-stem tails, the ka/mo/demo paradigm, and the purpose clauses to be given a position or dropped; forces the 35 -te dependents to be enumerated and ranked; forces the 140-item battery to be itemized; and forces the -eru separation and the kango 200-item window to be reconciled into one number. Every structural defect above surfaces on contact with a numbered list. That is exactly why the document does not have one, and exactly why it reads as “overly high level” despite its length.