Input rebalance

RE-DERIVED DAILY BRAID UNDER THE VOLUME CONCESSION

Headline finding before anything else: the trade was never cards-vs-input. With dwell-graded audio cards at ~6 s/review and ~20 s/new, even a 22/day new rate costs ~22 min/day at peak. The plan’s 12/20/32-minute input allocation was not caused by the card budget — it was caused by the coverage gate, which told the plan that input below ~95% lemma coverage was worthless and therefore not worth buying minutes for. The gate is the bug. Repeal the gate and roughly 50 free minutes/day appear at every phase without cutting a single card in the coverage-critical zone.

1. THE ACTUAL EVIDENCE ON COVERAGE THRESHOLDS FOR LISTENING

1a. Established (direct empirical work, listening as the criterion task)

Study	Design	Finding
van Zeeland & Schmitt (2013), Applied Linguistics 34(4):457–479	Nonword substitution in narrative listening at 90/95/98/100% coverage, n=40	90% coverage → mean comprehension 73.5%. 95% → 76%. Statistically similar. At 100%, 29/40 scored perfect, none below 7/10. They recommend 95% not because 90% fails but because comprehension at 95% is more consistent across listeners.
Bonk (2000)	Listening + coverage measurement across four texts	No coverage figure guaranteed comprehension. Some listeners had good comprehension at 80–89%. Below ~80%, good comprehension becomes very unlikely. The required coverage is a function of the comprehension criterion you demand — 98% is needed only if you demand ≥60% on a detail-level test.
Adolphs & Schmitt (2003), CANCODE	Coverage profile of spoken discourse	2,000 word families < 95%; 3,000 families ≈ 96%. The marginal return collapses after 2–3k in speech far faster than in writing.
Nation (2006)	Wellington spoken corpus	3,000 families + proper nouns ≈ 95% of unscripted spoken English; 6,000–7,000 ≈ 98%.
Stæhr (2009), SSLA 31(4)	115 advanced Danish EFL, listening test	Vocabulary breadth predicted 49% of listening variance. Big jump 3,000 (≈94% coverage) → 5,000 (≈98%).

1b. Assumed, not established — and the critic is right about this

The 98% figure is imported. Hu & Nation (2000) and Laufer & Ravenhorst-Kalovski (2010) are reading studies with reading criteria. Nation’s “6,000–7,000 families for spoken text” is not an empirical listening-comprehension result — it is the reading-derived 98% criterion applied to a spoken corpus frequency profile. The plan inherited a reading threshold wearing a listening costume.
There is no breakpoint in either modality. Schmitt, Jiang & Grabe (2011) found the coverage–comprehension relation is essentially linear with no step. “Threshold” is a decision rule about acceptable comprehension, not a property of language.
Stæhr’s 98% is the one listening result that supports the high figure — and it is the least transferable. It used an advanced academic listening test with detail questions on unfamiliar topics, single pass, no interaction. That is the maximum-difficulty corner of listening. It says nothing about conversation, where the listener can request repair.
Japanese-specific coverage curves for conversation are not published in any form I could locate. The CEJC (200 h, ~2.4M words, NINJAL, 2022) exists but I found no per-rank coverage table. The English figures are the only available proxy. My working assumption — marked as approximate and unverified — is that Japanese conversational coverage at a given lemma rank runs higher than English at low ranks (heavier functional-morpheme token share, extremely formulaic backchanneling) and lower at high ranks (kango compounding is productive, so the tail is fatter). Net: “~3,000 lemmas ≈ 95% of conversation” is a reasonable but untested import.

1c. The defensible lower threshold

~90% coverage of the specific text, for adequate gist of familiar-topic conversational or narrative speech. (van Zeeland & Schmitt: 73.5% mean comprehension.) ~80% is the floor below which comprehension collapses (Bonk).

And here is the move the original plan missed entirely:

Coverage is a property of a TEXT, not of a learner. 90% per-text coverage is reachable at N=800 by narrowing the topic and repeating the text. It does not require waiting for N=3,000.

Chang & Read (2006) is the direct empirical support: of four pre-listening supports, topic preparation was the most effective and vocabulary pre-teaching was the least effective, with repeated input close behind topic preparation. The plan’s core mechanism for making input comprehensible (pre-teach the words via cards) is the one that tested worst. Narrowing and repeating tested best.

Concretely, this is why 90% per-text coverage is cheap in Japanese. The following ~100 items plausibly carry 50–60% of conversational tokens under short-unit tokenization (approximate — extrapolated from BCCWJ short-unit distributions, not measured on CEJC):

Particles and functional morphemes (1–40)

no · 2. wa · 3. ga · 4. o · 5. ni · 6. to · 7. de · 8. mo · 9. kara · 10. made · 11. ka · 12. ne · 13. yo · 14. na · 15. sa · 16. -tte · 17. n (nominalizer) · 18. da · 19. desu · 20. -masu · 21. -nai · 22. -te · 23. -ta · 24. -tara · 25. kedo · 26. shi · 27. node · 28. noni · 29. dake · 30. gurai · 31. toka · 32. nanka · 33. yappari · 34. chotto · 35. moo · 36. mada · 37. zenzen · 38. sugoi · 39. hontoo · 40. soo

Backchannels and fillers (41–60) — near-zero in written corpora, enormous in CEJC-like conversation. This block alone is the strongest argument for the plan’s corpus choice. 41. un · 42. uun · 43. ee · 44. hai · 45. aa · 46. eeto · 47. ano · 48. sono · 49. maa · 50. nee · 51. hee · 52. fuun · 53. soo soo · 54. naruhodo · 55. desu ne · 56. soo desu ka · 57. soo nan da · 58. wakaru · 59. tashika ni · 60. dakara

Deictics, full ko-so-a-do grid (61–80) 61. kore · 62. sore · 63. are · 64. dore · 65. koko · 66. soko · 67. asoko · 68. doko · 69. kono · 70. sono · 71. ano · 72. dono · 73. konna · 74. sonna · 75. anna · 76. donna · 77. koo · 78. soo · 79. aa · 80. doo

Highest-frequency verbs, spoken forms (81–100) 81. suru · 82. naru · 83. aru · 84. iru · 85. iu · 86. omou · 87. iku · 88. kuru · 89. miru · 90. wakaru · 91. dekiru · 92. kangaeru · 93. taberu · 94. nomu · 95. kaeru · 96. kiku · 97. hanasu · 98. yaru · 99. morau · 100. ageru

Add ~200 narrow-topic content items (one Teppei episode’s worth) on top of these 100, and a single 4-minute episode about, say, breakfast is at 90% per-text coverage. At N ≈ 300. In week 3.

1d. What comprehensible-input theory says about input BELOW the threshold

Krashen’s model has no coverage threshold. His construct is comprehensibility achieved by any means — context, gesture, pictures, prior knowledge, redundancy, repetition. The coverage literature measures only the lexical route to comprehensibility and is silent on the rest. That is a real gap in the plan’s reasoning, but it is not a licence for unlimited optimism.

The honest position: below-threshold input does not do comprehension work, but it does other work that this learner desperately needs, and that no card can do.

Segmentation via statistical learning. Saffran, Aslin & Newport (1996): 8-month-olds extracted word boundaries from a synthesized syllable stream in two minutes, using transitional probabilities alone, with zero semantics. Adults do this too. Segmenting the Japanese speech stream is a distributional achievement, not a lexical one.
Phonetic category formation. High-variability phonetic training (Logan, Lively & Pisoni, 1991) improves category boundaries using non-comprehended stimuli — and requires many talkers, not many hours with one.
Rhythmic segmentation routine. Otake, Hatano, Cutler & Mehler (1993): Japanese listeners segment by mora. Installing a moraic segmentation routine in an English-L1 adult is exactly this plan’s foundational bet, and it is a perceptual achievement acquired from input, not from cards.
Prosodic and phrasal chunking, including where turn boundaries fall — inaudible in any card.

The hard limit: none of that happens ambiently. The decades-of-TV counterexample is real, and Schmidt’s noticing hypothesis is the reason. Statistical-learning studies use attended, structured input. So: “input below threshold is valuable” is true only with the qualifier “for perception, when attended, under a task.” That qualifier is the entire content of §4.

One finding that cuts the other way and partially defends the original plan: Vidal (2011) found reading produces greater incidental vocabulary gains than listening at every proficiency level, with the gap largest at low proficiency. This learner has no reading channel at all. So the input’s vocabulary yield will sit at the low end of what the English literature predicts. The volume argument survives — but its lexical component is weaker for an audio-only learner than the critic’s framing implies. This is why the revision below keeps a 22/day peak in the coverage-critical zone rather than collapsing to the critic’s flat 15 minutes of cards.

2. RE-ALLOCATED DAILY MINUTES

Planning baseline: 90–100 min/day. Bad-day fallback 60 min (cut in the order given below). Good-day 120 min (every surplus minute goes to Track C).

Three named input tracks, with different jobs and different gates:

Track C — Comprehension input. Per-text coverage ≥90%, achieved by narrowing and repeating. Grows comprehension. Mandatory output artifact.
Track P — Perceptual input. Coverage irrelevant. Native speed, multi-talker, multi-party. Grows segmentation, category boundaries, prosody. Mandatory countable tally.
Track Z — Ambient. Budgeted at zero. Allowed, never counted, never substituted.

Phase 0 — Weeks 1–2 (pre-card perception block) · 75 min

Activity	Min
Perception drill (mora count, vowel length, gemination, moraic N) — ≥6 talkers per contrast	30
Track P (native-speed audio, tally task only)	15
Track C (picture-supported absolute beginner)	20
Mora-timed echo production	10

Attended audio: 45 min/day. Original: ~0. This is the single largest free win in the whole revision — the original plan spent two weeks with the input tap fully closed, for no evidential reason. Perceptual input is not coverage-gated.

Phase 1 — Weeks 3–4 · cards open at 12/day · 90 min

Activity	Min
Cards (12 new + ~45 reviews)	9
Perception maintenance (3 days/wk, new talkers)	12
Track C	35
Track P	12
Shadowing	15
Slack / spillover to Track C	7

Attended audio: 62 min/day. Original: 12.

Phase 2 — Weeks 5–10 · 18/day · 90 min

Activity	Min
Cards (18 new + ~95 reviews)	16
Perception maintenance + accent listening	8
Track C (narrow listening: 1 topic × 6 episodes, then repeat)	38
Track P	13
Shadowing	15

Attended audio: 66 min/day.

Phase 3 — Weeks 11–20 · 22/day (peak card rate) · 95 min

Activity	Min
Cards (22 new + ~135 reviews)	22
Perception maintenance	4
Track C	38
Track P	16
Shadowing + TTS-prompted response drills	15

Attended audio: 69 min/day. Original: 20. Note the card rate goes above the old 20/day cap here — deliberately. This window is where each item buys the most coverage.

Phase 4 — Months 5–9 · 18/day · 100 min

Activity	Min
Cards (18 new + ~150 reviews)	22
Track C	34
Track P	20
Shadowing	8
Live conversation (2–3×/wk, amortized)	16

Attended audio: 78 min/day. Original: 32.

Phase 5 — Months 9–13 · 12/day · 100 min

Activity	Min
Cards (12 new + ~140 reviews)	18
Track C (now near-authentic material)	38
Track P (now mostly unscripted multi-party)	20
Live conversation	24

Attended audio: 82 min/day.

Phase 6 — Months 13–24 · 8/day · 100 min

Activity	Min
Cards (8 new + ~110 reviews)	14
Track C	32
Track P	20
Live conversation	34

Attended audio: 86 min/day.

What gets cut, in bad-day order

The up-front perception block’s length — from a long dedicated stretch to 2 weeks × 30 min, then a decaying maintenance thread (12 → 8 → 4 → 0 min). Justification: high-variability phonetic training plateaus; what generalizes is talker variety, not hours. Six voices per contrast beats sixty repetitions of one.
Explicit grammar study — deleted as a standing budget line. It becomes ~20 min, once, at each of five readiness triggers (§5, L9). Total lifetime cost ≈ 100 minutes.
Isolated production drilling — shadowing survives (it is input and output and it is a comprehension test); decontextualized pattern practice does not.
Card review of the tail — new-card rate tapers to 8/day from month 13 (§3).
On a 60-min day: cut Track P first, then shadowing, then cards to reviews-only. Never cut Track C. Track C is the one line that is doing the work the concession says is 80% of the outcome.

Attended-audio hours accumulated, at 90% adherence:

Milestone	Revised	Original
Month 9	~290 h	~100 h
Month 24	~870 h	~500 h

2.9× at month 9, 1.75× at month 24 (the gap narrows because the original uncaps at month 9).

3. WHERE THE 6,000 DATE MOVES

Do not cut cards flat. Cut them by the shape of the coverage curve. Item #400 buys enormous coverage; item #5,400 buys perhaps 0.02%. So: front-load into the coverage-critical zone, taper hard in the tail where input does the work anyway.

Revised ramp: 12 → 18 → 22 → 18 → 12 → 8 per day.

Cumulative, at 90% adherence:

Point	Cumulative N	Revised date	Original date
1,000 items	1,000	week 11	week 12.5
2,500 items	2,500	month 5.2	month 5.6
End wk 20	2,218	—	1,940
Month 9	4,372	—	~4,600
6,000 items	6,000	month 14–15	month 12–13
Month 24	~8,000	—	~9,000–12,500

The 6,000 date slips ~2 months. The 1,000 and 2,500 marks arrive EARLIER. That is the whole point of tapering rather than flattening. For comparison, the critic’s cruder version (flat 10/day after week 10) pushes 6,000 to month 22 — that is the floor, and it is worse than necessary.

Is the slip acceptable? Yes, for three reasons, and one caveat.

The coverage math says the tail barely matters for conversation. Items 3,000→6,000 move spoken coverage from roughly 95% to roughly 98% (Nation 2006, Adolphs & Schmitt 2003). Those 3,000 items buy ~3 percentage points. Items 1–3,000 buy ~95. Arriving at 3,000 two weeks early and at 6,000 two months late is a straightforwardly good trade.
The binding constraint at 6,000 is not pile size, it is retrieval latency under time pressure — and that is trained by input and conversation, not by adding items. A learner with 6,000 recognizable-in-isolation items and 500 hours of listening is worse in conversation than a learner with 4,500 items and 900 hours.
In the tail, the card stops being the teacher and becomes the guarantee. Peters & Webb and the audiovisual incidental-uptake literature show input teaches mid-frequency vocabulary directly, just inefficiently. Once N ≈ 3,000 and input is heavy, items 3,000–6,000 arrive from input; the card’s job narrows to guaranteeing the ones input hasn’t delivered yet.

Caveat, stated honestly: Vidal (2011) says the incidental channel is weaker for listening than reading, and this learner has only listening. So argument (3) is softer here than in an English-reading context. That is exactly why the taper bottoms at 8/day rather than at zero, and why the peak is 22 rather than 20.

Change the goal metric. “6,000 words by month X” was never the objective; it is a proxy that has now been shown to trade against the objective. Replace it with:

Month 24 target: sustain a 30-minute unprepared conversation with a non-accommodating native speaker on everyday topics, and follow a 3-person unscripted conversation at ≥60% gist.

N becomes an input to that, not a target in itself.

4. WHAT KIND OF INPUT, IN ORDER OF WHEN IT BECOMES USEFUL

Constraint that eliminates most “graded” material: no reading. Anything whose comprehensibility depends on subtitles, transcripts, or furigana is unusable. What survives is material where meaning is carried by voice alone, drawings, gesture, or already-known plot.

Tier 1 — usable from day 1 (weeks 1–4)

Comprehensible Japanese (YouTube, Yuki) — Complete Beginner playlist. Whiteboard drawings carry the meaning; zero English; deliberately slow. The best day-1 resource that exists for a no-reading learner.
Nihongo con Teppei for Beginners — 3–5 min episodes, 600+ of them, one narrow topic each, all-Japanese, heavy self-repetition, no visuals required. This is the narrow-listening workhorse and it is what Chang & Read’s “topic preparation” finding predicts should work best.
Your own TTS card audio, replayed as continuous streams — 10–15 items looped in their “X ga” frames. Free, given the pipeline. Counts as Track C.
Children’s songs / dooyoo — for mora timing only. Warning: singing distorts both pitch accent and mora length. Never use as an accent model.

Tier 2 — from ~week 6–8 (N ≈ 600–900)

Comprehensible Japanese, Beginner tier
Sakura Tips — short, slow monologue
Bitesize Japanese Podcast — very short, scripted, narrow
Japanese with Shun — daily-life monologue, moderate speed
Nihongo Switch — two hosts, semi-scripted

Tier 3 — from ~week 14–20 (N ≈ 1,500–2,500)

Yuyu no Nihongo — natural but deliberate; the key bridge resource
Learn Japanese with Noriko — clear, topical, long-form
Nihongo con Teppei (main series) — native-speed rambling monologue
Onomappu — native speed, strong visual and gestural support
Miku Real Japanese — natural speed, clear articulation

Tier 4 — from ~month 5–7 (N ≈ 3,000+)

Slice-of-life anime with mundane settings and adult register: Shirokuma Cafe, Aggretsuko, Hanasaku Iroha, Sakura-sou. The anime trap is specific and it is not about difficulty — it is about register. Shonen and fantasy teach yakuwarigo (role language): ore-sama, -de gozaru, archaic copulas, -zo/-ze, shouting, invented vocabulary. Reproducing it marks you as strange in a way you cannot hear. Filter by setting, not by level.
Japanese dubs of shows whose plot you already know. This is topic preparation — the highest-value listening support Chang & Read tested — available for free.
J-drama with mundane settings: Shinya Shokudoo, Kodoku no Gurume. Slow, repetitive frames, adult register.

Tier 5 — Track P from month 3, Track C from month 7–9. This is the actual target.

Terrace House — genuinely unscripted, mundane, multi-party, overlapping. The closest publicly available approximation to CEJC-style input.
Native-for-native podcasts: Yuru Gengogaku Radio, COTEN RADIO, Rebuild.fm. Two-person unscripted, with backchannels, interruptions, false starts, and self-repair — none of which appear in graded material.
Variety / talk shows and YouTube zatsudan streams.
Live conversation with a human tutor (iTalki, Cafetalk, HelloTalk), from month 5. The only input that adapts to you in real time and the only one that produces negotiation of meaning.

What makes input “attended” rather than ambient

Track C — comprehension input. All four required:

Per-text coverage ≥ ~90%. Operational test: after one pass, you can retell the gist in two English sentences. If you can’t, it is not Track C — either narrow further, repeat it, or reclassify it as Track P.
A mandatory output artifact: the 2-sentence English retell, plus 3 Japanese chunks you caught verbatim. Logged.
Blocks of 10–20 minutes, then switch. Attention decays and 40 unbroken minutes is 15 minutes of listening and 25 of drifting.
Narrow, then repeat: stay on one topic or speaker for 3–6 sessions; re-listen to the same episode 2–3× across 3 days. Repetition converts partial parses into full ones (Chang & Read).

Track P — perceptual input. Coverage is irrelevant, but a countable target is mandatory. Pick one per session:

Count every geminate in 60 seconds (kitte, sakka, itchi, matta, otto)
Mark every long vowel (obaasan, ojiisan, biiru, tooru, kookoosee)
Tally every -teru
Tally every -tte
Tally every n da / n desu
Tally every backchannel (un, hai, ee, aa, naruhodo, soo desu ne) and note who takes the turn immediately after
Shadow 5-second chunks simultaneously, not after
Count speakers and mark every overlap and interruption

If you cannot state the tally at the end, it was ambient, not attended. Do not log it.

Track Z — ambient. Budgeted at zero. The honest statement: background Japanese has a small nonzero effect on phonotactic habituation and a large, reliable effect on making you feel like you’re studying. Only the second one is dependable.

The trap, stated precisely: Input far above level is decoration when it is asked to do comprehension work, and productive when it is asked to do perception work. The same Terrace House episode is decoration at month 3 if you watch it “to understand,” and genuinely productive at month 3 if you listen to count turn-transitions and catch every -tte. The variable is the task, not the material. This is what lets the revision buy 3× the volume without buying 3× the wishful thinking.

5. WHAT SURVIVES, WHAT DEMOTES — PER LAW

#	Law	Verdict
L1	Perception block before card 1	HARD GATE SURVIVES, reshaped
L2	Accent integer on every card + “X ga” frame	SURVIVES UNTOUCHED
L3	Contractions as citation form	SURVIVES — PROMOTED to load-bearing
L4	Reduction findings generally	SURVIVES — PROMOTED
L5	Interference-based spacing	SURVIVES AS HARD
L6	Conversational-corpus ranking (not written, not CSJ)	SURVIVES — PROMOTED
L7	wago:kango ratio; kango collocation frames	Ratio → descriptive. Frames → HARD.
L8	Grammar gate at verb #56 + ≥4 godan per coda class	DEMOTED to soft preference
L9	Five stem keys at fixed weeks	Order survives. Calendar demoted.
L10	New-card ramp 10/15/20, 20 hard cap	RESHAPED to 12/18/22/18/12/8
L11	“Below 95% coverage, input is decoration”	REPEALED

L1 — Perception block. Hard gate survives. This is the one thing volume cannot fix. Japanese-L1 speakers after decades of English immersion still fail /r/–/l/ discrimination: exposure without perceptual attention does not move category boundaries. If mora timing, length, and moraic N are not installed before card 1, every card trains a wrong representation and the extra 200 hours of input are being poured into a broken parser. Changes: shorten the up-front block to 2 weeks × 30 min but make it recurrent (12 → 8 → 4 min/day maintenance through week 20), require ≥6 talkers per contrast (single-talker training does not generalize), and open Track P input on day 1 — the original plan’s zero-input opening weeks have no defense, since perceptual input is not coverage-gated. Drill set, real items:

Vowel length — obasan/obaasan · ojisan/ojiisan · kado/kaado · biru/biiru · toru/tooru · yuki/yuuki · sato/satoo · ie/iie · koko/kookoo · shujin/shuujin · byooin/biyooin Gemination — kite/kitte · saka/sakka · buka/bukka · ichi/itchi · mata/matta · oto/otto · kako/kakko · shite/shitte · asari/assari · isho/issho Moraic N — mina/minna · kinen/kin’en · tani/tan’i · konyaku/kon’yaku/konnyaku · senen/sen’en Accent, in the “X ga” frame (this is why L2 exists) — hashi ga (chopsticks / bridge / edge) · ame ga (rain / candy) · kaki ga (oyster / persimmon) · ima ga (now / living room)

L2 — Accent integer + “X ga” frame. Survives untouched. A schema decision with ~zero minute cost that never gated anything. Keep exactly. It becomes more valuable under heavy input, since input is where accent is consolidated and the card is where it is anchored. Demote only drilling of accent, never recording of it.

L3 — Contractions as citation form. Survives and is promoted from stylistic choice to load-bearing. Under 3× input volume this is the law that protects the entire investment: if the card says -te iru and the input says -teru, the input is unparseable and the extra 200 hours are wasted. The inventory, in romaji:

-te iru → -teru / -ten · 2. -te iru n desu → -ten da / -ten no · 3. -te shimau → -chau (past -chatta) · 4. -de shimau → -jau / -jatta · 5. -te oku → -toku · 6. -de oku → -doku · 7. -nakereba (naranai) → -nakya · 8. -nakute wa → -nakucha · 9. -te wa → -cha · 10. to iu → -tte · 11. to iu ka → tteka / tsuka · 12. no da / no desu → n da / n desu · 13. -te iku → -teku · 14. -ranai → -nnai (wakaranai → wakannai, tsumaranai → tsumannai) · 15. -ru no → -n no (suru no → sun no) · 16. sore wa → sorya · 17. kore wa → korya · 18. de wa → ja · 19. -te mo → -tatte (iku → ittatte) · 20. -reba → -rya (sureba → surya) · 21. mono → mon (da mono → damon) · 22. ra-nuki: mirareru → mireru, taberareru → tabereru

L4 — Reduction findings. Survives, promoted. Same argument as L3.

L5 — Interference-based spacing. Survives as hard. Tinkham (1993, 1997), Waring (1997), Nakata & Suzuki (2019) all converge: semantically or phonologically clustered presentation slows acquisition. This is a cost-free reordering with a real effect. Nothing in the volume concession touches it. It is the single best-supported piece of the ordering work and it should be the last thing anyone questions.

L6 — Conversational-corpus ranking. Survives, promoted. Under heavy input the ranking must match the input distribution or the cards stop pre-teaching what you are about to hear. And it is the only source for items 41–60 above — the backchannel and filler inventory, which is a large token share of real conversation and which written corpora and the CSJ (3/4 lecture monologue) almost entirely omit.

L7 — wago:kango. Split verdict. The 85:15 → 40:60 ratio is descriptive — it falls out of the frequency ranking. Do not enforce it against the ranking; that would be tail wagging dog. The 2–4 mora collocation frame on kango cards survives as hard, and gets more important under volume: kango homophony (kooen 公園/講演/後援, kikan 期間/機関/器官) is a purely audio-only problem, and the frame is the only disambiguator available to a learner who cannot see the kanji.

L8 — Grammar gate at verb #56. Demoted to soft preference. Pienemann cuts both ways. You cannot teach a stage early — but you also cannot withhold a stage the input is already delivering, and the learner meets -te forms in Teppei episode 3. Restated: the gate governs when you run the explicit induction session, never whether input may contain the form, and it must never delay the input schedule. If the condition arrives late, input proceeds regardless. The -nu fact survives untouched — shinu is the sole member of its coda class, so it is memorized alongside iku; that is a fact about the language, not a pedagogical choice.

L9 — Five stem keys. Order survives, calendar demoted. The dependency order (dictionary → i-stem → -nai stem → -te → e-stem) is defensible on frequency and morphological grounds. The week numbers are precisely what the teachability hypothesis says you cannot schedule. Replace each with an input-derived readiness trigger:

Stem key	Old gate	New trigger
Dictionary / classes	week 6	300 items carded, ≥25 verbs among them
i-stem	weeks 6–7	can catch -masu in a 60-s graded clip, ≥80%
-nai stem	week 8	can catch -nai / -nakya in a 60-s graded clip, ≥70%
-te form	weeks 9–10	can tally -te chunks in a 60-s native clip at ≥70%
e-stem / potential	week 13	can catch -eru potentials in a 60-s native clip, ≥60%

This makes the input do the scheduling, which is the only scheduler the teachability literature endorses.

L10 — Card ramp. Reshaped to 12 / 18 / 22 / 18 / 12 / 8, derived from marginal coverage per item rather than from a flat cap. The peak deliberately exceeds the old 20/day ceiling.

L11 — “Below 95% coverage, input is decoration.” Repealed, not demoted. Replaced by: (a) Track C requires ≥90% coverage of the specific text, achieved by narrowing and repeating, not by waiting for N; (b) coverage is irrelevant for Track P; (c) the only genuine decoration is below ~80% per-text while attempting a comprehension task — the Bonk floor.

6. HONEST EXPECTED OUTCOME

Calibration note. FSI rates Japanese Category IV: ~2,200 classroom hours plus roughly equal self-study, ~4,400 h to ILR 3. But that figure includes kanji and full literacy, plausibly 30–45% of Japanese-specific effort. This learner is not doing “FSI minus reading time” — they are doing FSI with the single most expensive subsystem deleted. A spoken-only trajectory should be materially faster in wall-clock hours than the FSI number implies. The original plan underexploited this. All percentages below are calibrated estimates, not predictions with error bars — treat them as falsifiable checkpoints, not forecasts.

Month 9

	Revised (~290 h audio, ~85 h cards/perception, N≈4,400)	Original (~100 h audio, N≈4,600)
Graded/scripted monologue, familiar topic	85–95% gist	50–65%
Patient native, 1-on-1, daily topics	70–80%, sustain 15–20 min	35–50%, sustain 5 min
Native-speed unscripted 2-person podcast	30–45% (topic yes, detail no)	15–25%
3+ person overlapping conversation	20–30%	10–15%
Speaking	CEFR A2+/B1−, ILR 1/1+	A2, slower retrieval
Isolated-card accuracy	~88%	~93%

The original’s specific failure mode at month 9 is nameable: items scoring ~95% in isolation fail at ~40–50% in stream, because they were never encountered under coarticulation, reduction, or native speech rate. That is the classic audio-Anki gap, and 100 hours of listening is not enough to close it. The critic is right at month 9, and the margin is large.

The revision’s one clear win outside comprehension: mora timing and length contrasts should be near-native-listener-acceptable, because L1 was kept as a hard gate and drilled with multi-talker variability. Accent will be roughly right on carded nouns, unreliable on verbs and compounds.

Month 24

	Revised (~870 h audio, ~1,020–1,100 h total, N≈8,000)	Original (~500 h audio, N≈9,000–12,500)
Graded/clear monologue	~98% — solved	~90%
Native-speed unscripted 1-on-1, familiar topic	80–90%	60–70%
Unscripted 2-person podcast, unfamiliar topic	60–75%	45–60%
Multi-party overlapping	50–65%	30–45%
Speaking	B1+/B2−, ILR 2 — sustains unplanned conversation with a non-accommodating native	B1−, ILR 1+/2−
Recognition vocabulary breadth	lower	higher — the original’s one real win

Named holes at month 24 in both versions, since the conversational-corpus ranking underweights them: keigo and business register; abstract and technical vocabulary; and everything requiring literacy — the deliberate trade.

The honest costs of the revision

Vocabulary breadth is lower at every single date. Negligible at month 9 (~4,400 vs ~4,600); real at month 24 (~8,000 vs ~9,000–12,500). If the goal ever reverts to a word count, the revision loses.
The audio-only incidental channel is weak (Vidal 2011). The critic’s “3–4×” volume argument was built on literature where readers exist; this learner’s input yields less vocabulary per hour than that literature implies. The direction of the effect is unchanged; the magnitude is smaller.
Cards are self-verifying. Input is not. This is the largest execution risk and it did not exist in the card-heavy plan. Seventy-five minutes of nominal listening is trivially faked into 20 minutes of listening and 55 of drifting; 22 minutes of cards is not fakeable. This is the entire reason Track C and Track P both carry a mandatory output artifact. Without the retell log and the tally log, the revised plan is strictly worse than the original — it will produce fewer cards and no more comprehension.
Outcome variance is wider. The card plan has a narrow, mediocre distribution. The revised plan has a wider one with a better mean.

Falsifiable checkpoints — correct the plan against these, not against the estimates above

When	Test	Failure → action
Week 8	Unheard Teppei for Beginners episode → correct 2-sentence gist retell	Per-text coverage too low. Narrow harder; repeat more.
Week 20	Tally -teru in a 60-s native clip, ≥70% accuracy	Reduction ledger isn’t landing. Move 8 min from Track C to Track P.
Month 5	10-min tutor conversation on a prepared topic, zero English	Output is under-fed. Move 10 min from Track P to shadowing/production.
Month 9	Unheard Yuyu no Nihongo ≥70% gist AND unscripted 2-person native podcast ≥30% gist	If the second is <15%, the “input” was ambient. Audit the tally logs.
Month 13	30-min unprepared tutor conversation, <5 English words	—
Month 24	Unheard Terrace House episode, 3-person conversation, ≥60% gist	—

Bottom line

The revised braid wins at month 9 by a large margin and at month 24 by a moderate one, and it wins on precisely the metric that was declared the goal. The original wins on item count, which was never the goal. The ordering work is not discarded — L1 through L7 survive intact or strengthened, and L3/L4/L6 become more important under volume, because the reduced, conversational forms are what the extra 200 hours actually consist of. What dies is one law: the imported 95% gate. That single repeal buys almost the entire volume increase, at the cost of two months on a date that does not matter.

Sources: