Tang poetry · nine readers · one anthology, every test counted

The Tone That Went Missing

Tang regulated verse sets level tones against oblique ones by rule. Score the 213 quatrains and octaves of Three Hundred Tang Poems with the rhyme book of 1008 and the rules hold on 97% of tests; read them in Cantonese and they hold on 96%; read them in Mandarin and more than one test in ten fails. Nearly every failure Mandarin adds falls on the entering tone, the syllables that once ended in -p, -t or -k. Six more readers, predicted before the count: Hán-Việt and Hokkien keep the rules as the rhyme book does, Korean or Japanese readings mend Mandarin, and Hakka, read from its dictionary, breaks them almost as often as Mandarin, for a different reason.

Seven level tones in a row

This is Du Fu's poem on hearing that the government armies had retaken Henan and Hebei, 聞官軍收河南河北. Each square is one syllable, shaded by its tone class: level (平) or oblique (仄). The buttons swap the reader. The first is the Guangyun, a rhyme dictionary compiled in 1007 and 1008, which files every character under one of the four tones (some under more than one). The other two are how a Cantonese speaker and a Mandarin speaker read the same characters today. The second row holds six more readers, added on 7 November 2026 and explained below. The outlined squares are the positions the rules care about, and a red border marks a rule the reading breaks.

Under Mandarin, the seventh line, 即從巴峽穿巫峽, comes out as seven level tones. Three of its syllables (即, 峽, 峽) are in the entering tone in the rhyme book, and Mandarin now says them with a rising tone, which counts as level. In the rhyme book the line passes every test that can be scored on it; the rule was built for a tone Mandarin no longer has. The Mandarin reading breaks three more rules on 看, “look”, in line 3, which both modern readers give an oblique tone (kàn in Mandarin, hon3 in Cantonese). The rhyme book lists 看 in both the level and the departing tone, so those three are left out of the count below.

What the rules ask

The four tones of Middle Chinese were called level, rising, departing and entering. Poets sorted them into two: level (平) on one side, and the other three, together called oblique (仄), on the other. Regulated verse (近體詩, the quatrain 絕句 and the eight-line 律詩) is defined by how those two alternate. The mnemonic says “一三五不論，二四六分明”: the first, third and fifth syllables do not matter, the second, fourth and sixth must be clear. The rules scored here are the textbook ones, as the Chinese Wikipedia states them.

• R1, within the line: the second and fourth syllables differ, and in a seven-syllable line the fourth and sixth differ too.
• R2, opposition (相對): the two lines of a couplet have opposite tones in their second syllable.
• R3, adhesion (相黏): the first line of each later couplet has the same tone in its second syllable as the line before it.
• R4, rhyme: the even-numbered lines end in a level tone.
• R5, the unrhymed lines: the odd-numbered lines after the first end in an oblique tone.

Those five rules fix fewer syllables than they seem to. Pick the tone of the first line's second syllable and whether the first line rhymes, and they settle the rest of what they touch in a quatrain. Everything dashed below they leave free.

All 213 quatrains and octaves

Three Hundred Tang Poems was compiled around 1763 by Sun Zhu and sorts its poems by form. Its four sections of five- and seven-syllable quatrains and octaves hold 213 poems, scored here as regulated verse, and every one of them fits the scoring. The old-style sections (古詩) are the control: many of the same poets, the same dynasty, and no tone rule. Some tests cannot be scored by every reader: a character listed in both a level and an oblique tone in 1008 (看, “look”, is one) could be either, so the table counts only the tests that all three readers can score. That leaves 3,396 of the 4,354 regulated tests.

In the regulated poems the rhyme book and Cantonese agree to within two points on every rule. Mandarin loses between eight and ten points on every rule except the rhyme, where it loses nothing. In the old-style poems the adhesion rule holds 54.9% of the time, close to what a coin would give: the rules are real, and they belong to the form, not to the language. One thing the control did not show was indifference to opposition: old-style couplets still set their second syllables against each other 79.0% of the time.

Where Mandarin breaks, and why

Of the tests that pass in the rhyme book and fail in Mandarin, 246 of 253 involve a syllable whose tone in 1008 was entering, and in 245 of them that syllable is the one Mandarin moved into the level class. (In the other, 莫 kept an oblique tone and 場 beside it changed.) In the regulated poems 984 syllables have only an entering reading. The Mandarin reader gives 368 of them a level tone and 615 an oblique one (and leaves one toneless). The Cantonese reader gives all 984 an oblique reading, because every one of them still ends in -p, -t or -k.

That the entering tone makes regulated verse misbehave in Mandarin is not news. The Chinese Wikipedia's article on the tone classes gives the standard example, the first line of another poem by Du Fu:

“由於現代标准官話裡沒有入聲，因此用現代标准官話唸詩，可能會出現「平仄不調」的情況。在以上例子中，雖然「國」字屬於入聲，但在現代标准官話裡卻是陽平聲（第二聲）。”

zh.wikipedia.org, 平仄 (fetched 2026-09-29). In English: because modern standard Mandarin has no entering tone, reading poems in it can make the tones come out wrong; 國 is entering, but in Mandarin it has the second tone, which is level.

Their example falls on the first syllable of the line, one of the positions the mnemonic leaves free, so it breaks nothing the rules check. What this page adds is the count at the positions that are checked, over one whole anthology, with the control and the second modern reader beside it.

What the rhyme book does not explain

The rules do not hold on every test even in 1008, and each kind of miss is worth seeing.

• Quatrains that rhyme on oblique tones. Nine poems in the anthology's five-syllable quatrain section rhyme on oblique tones throughout, which makes 18 of the 22 rhyme failures. Among them are Wang Wei's 鹿柴 and 竹裏館, Meng Haoran's 春曉, and Liu Zongyuan's 江雪, which rhymes on entering-tone words (絕, 滅, 雪). They sit in the quatrain section, but the tradition does not count every quatrain as regulated: the Chinese Wikipedia calls the older kind 古絕 and describes it as freer in form, without strict metrical rules, and often rhymed on oblique tones (“形式較為自由，並無嚴格的格律限制，常有押仄聲韻的作品”). The pre-registration scored every poem in the quatrain sections as regulated, so they stay in the count.
• A copying error. Three rhyme failures are the text, not the poet. The digital edition used here prints 幹 (“trunk”, departing tone) where the poems have 乾 (“dry”, level) in 逢入京使, 月夜 and 無題 (相見時難); Wikisource has 乾 in all three. The same error turns the place name 長干 into 長幹 in two other poems. Put those five characters back and the rhyme rule goes from 561 to 564 of 583. The table above keeps the text as it came, because that is what was pre-registered.
• One gap in the data. The last rhyme failure is 漫, in Meng Haoran's 早寒有懷, which the rhyme-book data used here lists only in the departing tone.
• A licensed shape. Of the 67 within-line failures, 26 end in one shape: an odd-numbered line whose last three syllables run oblique, level, oblique, the ending of what the tradition allows as the “first special form”, (仄仄)平平仄平仄. Only 21 of the 26 fit that form as the source describes it: in 江雪 the first line rhymes, and four begin with an oblique syllable where the form wants a level one. Nine more would end in the shape if an ambiguous syllable is read oblique. The rest include old-style quatrains, Li Bai's 靜夜思, and the line 黃鶴一去不復返 in Cui Hao's 登黃鶴樓.

Poems that break no rule on the tests all three readers can score: 151 of 213 in the rhyme book, 136 in Cantonese, 65 in Mandarin.

What was predicted, before the count

The predictions were committed at 10:26 UTC on 29 September 2026, before any rate was computed, and are scored here as written.

• P1 · PARTLY FAILED. Regulated verse in the rhyme book: R1, R2 and R3 each at least 95% (held: 95.5, 98.7, 96.7), and R4 at least 98% (failed: 96.2, for the reasons above).
• P2 · PARTLY FAILED. Old-style verse in the rhyme book: adhesion at most 70% (held: 54.9) and opposition at most 75% (failed: 79.0).
• P3 · HELD. Mandarin between 5 and 20 points below the rhyme book on R1, R2 and R3: 8.4, 9.2 and 8.5 points.
• P4 · HELD. Cantonese within 3 points of the rhyme book on the same three: 0.6, 0.6 and 1.1.
• P5 · HELD. At least 80% of the Mandarin-only failures involve an entering-tone syllable: 246 of 253, 97.2%. The other seven: 場 read chǎng (four tests), 值 read zhí, which the data lists only in the departing tone (two), and 俱 read jù, a level tone that became oblique (one).

Six more readers, predicted before the count

Added 7 November 2026. The question this page left open was whether other readers who kept the old stop endings would hear the rules as Cantonese does. Six more readers were added, and nine predictions about them were committed before any of them was scored (the pre-registration, landed at 00:47 UTC on 7 November as commit 3816e7e7; the results). The corpus, the five rules and the tests are the ones above, unchanged.

• Hán-Việt, the Sino-Vietnamese readings of the characters. Vietnamese verse sorts the six Vietnamese tones into two classes named with the same two characters, bằng (平: ngang and huyền) and trắc (仄: sắc, hỏi, ngã and nặng), and the earliest Vietnamese poems that survive are Tang regulated verse, written in Chinese. Sino-Vietnamese kept the stop endings, and a syllable that ends in one carries only sắc or nặng.
• Hokkien and Hakka, two Chinese languages of the south-east. Hokkien and most Hakka varieties keep -p, -t and -k. Many characters in both have a literary reading (文讀) beside a colloquial one (白讀), and in Taiwan's traditional schooling classical texts were recited in the literary one; where the dictionary names the literary reading, that is the one used here. Hakka was the reader expected to slip: in most Hakka varieties some common characters that had a rising tone and a voiced initial are read in the level tone 陰平, and in Taiwan's Sixian these are often the colloquial readings, the literary ones staying rising or departing.
• Mandarin, mended with Korean or Japanese. Sino-Japanese readings preserve no trace of the Chinese tones, and Korean, which kept them for a while, has since lost them in all but a few dialects, so neither can read these rules alone. Both kept the endings. Korean keeps -k and -p, and its -l stands for the old -t (一 is 일, il). Japanese spells them -ku or -ki, -tsu or -chi, and, for -p, -fu, which the pronunciation lost and the spelling kept until the modern kana usage of 1946: 十 was borrowed as zipu and is now jū, written じふ before 1946 and じゅう after. These two readers do one thing: where Mandarin says level and every Korean (or Japanese) reading of the character ends in a stop, the syllable becomes oblique. Japanese is tried twice, in the old spelling and in the new.

Hán-Việt, Hokkien and Hakka are dictionaries, not speakers: each takes the readings its source lists for a character (for Hokkien and Hakka, the literary ones where the source names them), and the mended readers take Mandarin's reading and every Korean or Japanese one. A character whose readings fall in both classes is left out of the count, as in the rhyme book. Each is scored on the tests that it, the rhyme book, Cantonese and Mandarin can all score, so every comparison below is like for like.

Four of the six hear the rules as the rhyme book does, or within a few tests of it. On the tests each can score, Hán-Việt falls short of the rhyme book by 0.0 points on R1 and R3 and passes two opposition tests more than the rhyme book on R2, and Hokkien falls short by 0.0, 0.0 and 0.0: of its 3,121 tests, not one that passes in the rhyme book fails in Hokkien. Mandarin mended with Korean falls short by 0.1, 0.2 and 0.6 points on R1, R2 and R3, and mended with Japanese in the old spelling by 0.1, −0.2 and 0.3. What the mending leaves are a few Mandarin readings it cannot reach: 場 read chǎng and 俱 read jù, which have nothing to do with the entering tone, and 值 read zhí, which reflects an entering reading the rhyme book of 1008 does not list (a later rhyme book, the Jiyun, does) and whose Korean reading, 치, keeps no ending.

Hakka was predicted to slip, by at least a point more than Cantonese and by less than half as much as Mandarin. It slipped much further. On its tests it falls 7.5 points short on R1, 10.2 on R2 and 9.8 on R3, against Mandarin's 8.4, 10.7 and 8.2 on the same tests: about as far, and on adhesion further. The cause is the one predicted. Of the 193 tests that pass in the rhyme book and fail in Hakka, 173 involve a syllable that had a rising tone and a voiced initial and that Hakka reads in a level tone. Hakka reads none of the old entering syllables only as level; it breaks the rules by having moved part of another tone into the level class.

These are the readings the dictionary gives, and for Hakka that matters more than for the others. English Wiktionary names a literary reading for only 59 of the 1,987 corpus characters it gives Sixian readings for, so most of the readings here are the everyday ones. A paper by 呂嵩雁 among the Taiwan Ministry of Education's Hakka dictionary resources lists rising or departing literary readings beside the level everyday ones of 有, 語, 滿, 下, 近, 上 and 淡; a reciter who uses literary readings would avoid some of these breaks, and how many, this page does not measure. The Meixian readings, a second Hakka variety, slip the same way: 6.0, 8.6 and 7.4 points.

Japanese in the new spelling slips where it was predicted to, and only there. Of the old entering syllables that ended in -p, the old spelling marks 115 of the 117 it reads one way as entering and the new spelling none of them; on -t and -k the two agree. So Mandarin mended with the new spelling falls 1.1 points short on R1, against 0.1 with the old, a difference of 1.04 points against the 1 predicted. The syllable that does more of the breaking than any other is 峽, the gorge in Du Fu's seventh line (10 tests): its go-on and kan-on readings were written げふ and かふ before 1946 and are now ぎょう and こう. In the first instrument, Mandarin mended with the old spelling turns both 峽 back to oblique; with the new spelling they stay level, and the line breaks the same tests it breaks in Mandarin.

What was predicted, before the count

Landed on main at 00:47 UTC on 7 November 2026, before any of these readers was scored, and scored here as written.

• P6 · HELD. Hán-Việt within 2 points of the rhyme book on R1, R2 and R3: 0.0, −0.4 and 0.0.
• P7 · HELD. Hokkien within 2 points of the rhyme book on the same three: 0.0, 0.0 and 0.0.
• P8 · FAILED. Hakka on R1 at least 1 point further short than Cantonese (held: 7.5 against 0.6) and less than half as far short as Mandarin (failed: 7.5 against 8.4).
• P9 · HELD. At least half of the tests Hakka fails and the rhyme book passes involve an old rising syllable with a voiced initial that Hakka reads level: 173 of 193.
• P10 · HELD. Mandarin mended with Korean within 2 points on R1, R2 and R3: 0.1, 0.2 and 0.6.
• P11 · HELD. Mandarin mended with Japanese in the old spelling within 2 points: 0.1, −0.2 and 0.3.
• P12 · HELD. With the new spelling, at least 1 point further short on R1 than with the old: 1.1 against 0.1, a difference of 1.04.
• P13 · HELD. Hán-Việt, Hokkien, Hakka, Korean and Japanese in the old spelling each keep at least 95% of the old entering syllables they read (99.9, 100, 100, 95.7 and 97.4%), and the new spelling at least 5 points fewer (85.7%).
• P14 · HELD. Korean and Japanese mark fewer than 1% of the syllables with no entering reading in the rhyme book as entering: none of about 6,900, in either spelling.

What these readers are not. Hán-Việt, Hokkien and Hakka are dictionaries, not speakers: they never make the choice a reciter makes between two readings, so they never get it wrong and never get credit for getting it right. Hán-Việt is one compiler's list, built from the dictionaries named in the sources below. Of its three readings that break a rule here, 婢 tỳ is the standard Hán-Việt reading; for 幡 the list gives only phướn, which the online Hán Nôm dictionary at hvdic.thivien.net, one of the list's own sources, gives as a Nôm reading beside the Hán-Việt phan and phiên; and for 榴 it keeps only lựu where that dictionary gives lưu and lựu. Hakka, Hokkien, Korean and Japanese are read from English Wiktionary, a dictionary written by volunteers. The mended readers start from the Mandarin reader's own choice of reading, and Korean or Japanese only correct it; they are not a way anyone reads a Tang poem, and measure only what a Mandarin reader gains by knowing a character's Korean or Japanese reading.

What this does not show

The rhyme book is not a recording. The Guangyun is a revision of the Qieyun of 601 and records which characters rhymed and which tone each was filed under; it does not say how the tones sounded, and their pitch is now hard to determine. The three readers are software: the rhyme-book reader is the qieyun package's Guangyun data, and the two modern readers are dictionaries that choose one reading for each character from its context. They make mistakes, and some show here: the Mandarin reader reads 乾 “dry” as qián, and where the Cantonese reader and the rhyme book disagree on a character, the commonest case is 斜 read ce3, which may be the software's reading rather than any speaker's. A character the rhyme book lists in both classes is never scored, which removes the hard cases along with the easy ones: 看 in Du Fu's third line is one, and both modern readers read it oblique where the rule needs it level. Thirteen characters could not be found in the rhyme-book data even through their variant forms and are not scored. The anthology here is one digital edition of one anthology.

Sources

1. 唐詩三百首, in chinese-poetry/chinese-poetry, 蒙学/tangshisanbaishou.json, commit b8594f81 (MIT licence). GitHub.
2. nk2028, qieyun 0.13.5 (CC0), a JavaScript library for the Qieyun phonological system; its data project describes its Guangyun table as the 澤存堂 edition with corrections. npm, data.
3. CanCLID, to-jyutping 3.1.1 (BSD-2-Clause). npm. And pinyin-pro 3.29.4 (MIT). npm.
4. Unicode Consortium, Unihan database, Unihan_Variants.txt, Unicode 18.0.0, for the variant forms. unicode.org.
5. Wikipedia (Chinese): 近體詩, 律诗 (the special forms) and 平仄 (the Mandarin example), fetched 2026-09-29. 近體詩, 律诗, 平仄.
6. Wikipedia (English): Four tones (Middle Chinese), Checked tone, Guangyun, Three Hundred Tang Poems, fetched 2026-09-29.
7. Wikisource (Chinese): 月夜 (杜甫), 逢入京使, 無題 (相見時難別亦難), 長干行二首, 長干曲其二, for the five corrected characters, fetched 2026-09-29.
8. English Wiktionary, the page of every character in the corpus, fetched 2026-11-07: Hakka (Sixian, in Pha̍k-fa-sṳ, and Meixian), Hokkien (in Pe̍h-ōe-jī, with its notes on literary and vernacular readings), Korean and Japanese (go-on and kan-on, in the old and the new spelling) readings. Wiktionary contributors, CC BY-SA 4.0; for that reason more-readers.json is offered under the same licence. An example page (峽).
9. Phong Phan, hanviet-pinyin-wordlist, commit b9923df9 (MIT licence), the Hán-Việt reading of each character, compiled from Thiều Chửu's Hán Việt tự điển (1942), Trần Văn Chánh's Từ điển Hán Việt (1999), Nguyễn Quốc Hùng's Hán Việt tân từ điển (1975), Đặng Thế Kiệt's Hán Việt Từ Điển Trích Dẫn, hvdic.thivien.net and CC-CEDICT. GitHub.
10. Unicode Consortium, Unihan database, Unihan_Readings.txt, Unicode 18.0.0: kHangul, kJapaneseOn and kTang (Hugh M. Stimson, T'ang Poetic Vocabulary, 1976), for the cross-checks. UAX #38.
11. Wikipedia (English), fetched 2026-11-07: Vietnamese poetry, Vietnamese phonology, Checked tone, Varieties of Chinese, Hokkien, Literary and colloquial readings, Sino-Xenic vocabularies, Sino-Japanese vocabulary. Vietnamese poetry, Checked tone, Sino-Xenic vocabularies.
12. Wikipedia (Chinese), fetched 2026-11-07: 客家話 (the rising tone read level) and 入聲 (一 in Korean). 客家話, 入聲.
13. 呂嵩雁, 客家語的文白異讀, among the Taiwan Ministry of Education's Hakka dictionary resources, fetched 2026-11-07. PDF.
14. Agency for Cultural Affairs (Japan), 現代仮名遣い: the preface and the table of older spellings, fetched 2026-11-07. preface, table.
15. 漢典 (zdic.net), 值, for its Jiyun reading; and the Hán Nôm dictionary at hvdic.thivien.net, 婢, 幡 and 榴, fetched 2026-11-07. 值, 婢, 幡, 榴.