Logos52
journal / 2026 06 21 hanviet bridge deep dive

The Hán-Việt bridge — deep dive (linguistics, learning, and how it builds into tsumugu-ed)

The Hán-Việt bridge — deep dive (linguistics, learning, and how it builds into tsumugu-ed)

Verdict: Sound is one of six channels of the Hán-Việt bridge, and the weakest of them for a learner. The payload that helps a Vietnamese student of Mandarin is lexical-morphemic: they already command ~3,000 Sino-Vietnamese morphemes as living vocabulary and can parse 學生 as học+sinh before meeting xuéshēng. Our current bridge (a hanViet string on 2,659/2,662 chars plus one example word on 940) ships only the thin sound version. The recommended build is the MC-Pivot Bridge: a Phase 1 with zero schema change and zero authoring (a third reading mode plus a derived entering-tone diagnostic), then each schema change earned behind a usage signal. The cost: one number the whole tone feature rests on does not exist publicly and must be measured locally first.

This was a thorough pass — a 68-agent, ~2.5M-token research-and-design workflow, web-verified, grounded in the live repo. 56 high-risk claims were adversarially checked; 6 were corrected and the corrections are folded in below.

What the bridge is

Hán-Việt is not derived from Mandarin and Mandarin is not derived from Hán-Việt. Both descend from Middle Chinese (MC), the language of the Qieyun rhyme dictionary (601 CE). Vietnam borrowed its systematic reading layer from the late-Tang (8th–9th c.) Chang’an literary standard during the ~1,000 years of Chinese administration (111 BCE–938 CE). The model: Hán-Việt ← Middle Chinese → Mandarin. Where MC made a distinction Mandarin merged away but Vietnamese kept, the Hán-Việt reading is a window onto the MC syllable and a partial predictor of the pinyin.

The “60–70% of Vietnamese is Sino-Vietnamese” figure is register-bound: ~⅓ averaged across all speech, ~40% by dictionary count, 60–70% in formal/technical/written registers only. The honest analogy is English’s Latin/French layer — a minority of everyday speech, a majority of academic prose.

The three strata (the load-bearing distinction)

The same character sits in Vietnamese at three depths, and conflating them is the main way a bridge mis-teaches.

  1. Old / Early Sino-Vietnamese (~400 words) — pre-Tang, fully nativized, felt as native. Safe exemplars: buồng (房), mùa (務), mùi (味), vốn (本), việc (役), đũa (箸).
  2. Sino-Vietnamese proper (~3,000 morphemes) — the rule-governed Tang literary layer. The predictable bridge; pairs morpheme-for-morpheme with Mandarin. The correspondence rules below apply only to this stratum.
  3. Vietnamized SV (越化) — readings that took a literary form then drifted in tone/sense.

The sound bridge

Initials — the near-deterministic rules

Hán-Việt→ MandarinMC sourceExamples
nh-r-palatal nasal 日nhân/rén 人, nhật/rì 日, nhập/rù 入
ng-/ngh-∅ (vowel)velar nasal 疑ngũ/wǔ 五, nghiệp/yè 業
ph-f-labiodental 非敷奉pháp/fǎ 法, phi/fēi 飛
h-x-/h-曉/匣học/xué 學, hải/hǎi 海
đ-d-dental stop 端đông/dōng 東
th-t-透thiên/tiān 天
v-w-/y-glide 云vương/wáng 王, vũ/yǔ 雨

Traps: MC dental s- surfaces as SV t-/th- (心 = tâm, 三 = tam), and SV velars stay velar while Mandarin palatalizes to j/q/x before front vowels (家 gia/jiā, 見 kiến/jiàn).

Codas — the goldmine

Vietnamese kept all six MC codas; Mandarin lost all three stops and merged -m into -n.

SV coda→ Mandarin
-p / -t / -c / -chopen syllable (was MC entering tone)
-m-n (三 tam→sān, 心 tâm→xīn, 金 kim→jīn)
-n-n
-ng / -nh-ng

Decision rule: stop-final SV ⇒ open Mandarin syllable; -m ⇒ -n; -n/-ng ⇒ same nasal. This rarely fails — the strongest single predictor in the bridge.

Tones — and the entering-tone payoff

Both tone systems descend from MC’s 4 tones × 2 registers (yin/yang, set by initial voicing) = 8 categories. The strong arrows: huyền → Mandarin T2; hỏi → T3; open-syllable sắc/nặng → T4. ngang splits T1/T2 on register (unrecoverable from the SV tone alone).

The payoff: a sắc/nặng tone on a stop-final syllable (-p/-t/-c/-ch) flags a 入聲 character — exactly the words whose Mandarin tone scattered across all four tones (入派三聲). In Vietnamese they are perfectly systematic; in Mandarin they are the “irregular” ones to memorize. Hán-Việt identifies which Mandarin tones are unpredictable, turning Mandarin’s chaos into a named category.

The honest limit: there is no peer-reviewed SV→Mandarin per-tone accuracy figure. The “81%” sometimes quoted is a Beijing-dialect number, not Sino-Vietnamese. Tone, exact vowel, and the doublet/irregular residue are where the rule stops and a lookup begins.

Beyond sound — the real value

  1. Morphemic recognition (the biggest asset). ~3,000 SV morphemes are already bound, productive vocabulary: 國 quốc (quốc gia, quốc tế, Trung Quốc), 學 học (học sinh, đại học, khoa học), 家 gia (gia đình, chuyên gia). The learner parses 科學 = khoa+học = “science” compositionally; a monolingual English learner memorizes an opaque blob.
  2. False friends (the brake you must ship). 困難 kùnnán “difficulty” → VN khốn nạn “vile, despicable”; 緊張 jǐnzhāng “nervous” → VN khẩn trương “urgent, hurry up”; 書院 → VN thư viện “library” but Mandarin shūyuàn “academy”. The morpheme match generates a hypothesis; the per-compound gloss is the truth.
  3. Word-order. SV compounds keep Chinese modifier-head order, contrasting with native head-initial Vietnamese: 白馬 bạch mã (white-horse) vs ngựa trắng; 飛機 phi cơ vs máy bay. A built-in minimal pair for the hardest structural difference.
  4. Etymological doublets (what makes a character felt). One etymon, two reflexes — an early native-feeling one plus the literary bridge form: 力 sức/lực, 鏡 gương/kính, 本 vốn/bản, 味 mùi/vị, 房 buồng/phòng. Learning 力 = lực is discovering that sức, strength felt in the arms, is the same morpheme.
  5. Wasei-kango caveat (和製漢語). The most “advanced” abstract terms — kinh tế 經濟, xã hội 社會, khoa học 科學 — are Meiji-Japanese coinages that flowed into both Chinese and Vietnamese ~1900. The trilateral match is reliable; framing them as ancient Chinese cognates is false.
  6. Chữ Nôm / orthographic. Surfacing the source character (học → 學) converts a sound fact into the graph the learner needs anyway. Guard: Nôm coined Vietnamese-only characters (𠀧 ba, 𤾓 trăm) that point to no usable Chinese reading — the tool must tag chữ Hán (transferable) vs Nôm-coined (Vietnamese-only).

Pedagogy, honestly

The cognate facilitation effect is largely automatic: learners acquire cognates better than non-cognates without instruction, and awareness-raising adds little to acquisition (it mainly helps false-friend avoidance). Design consequence: surface the correspondence and let it work; heavy explanatory scaffolding buys little.

No VI-L1 → L2-Mandarin cognate study with effect sizes appears to exist. General cognate gains are well attested; the Vietnamese-specific gain is plausible and unstudied, so no efficacy claims go user-facing. Pronunciation fossilization is a real risk — pinyin stays the default reading, Hán-Việt is opt-in, and the bridge always shows the Mandarin tone as truth. Central/Southern Vietnamese merge hỏi and ngã, so those tone heuristics break for a large share of users.

Current state in tsumugu-ed (measured)

hanViet on 2,659/2,662 chars. hanVietExample {word,han,glossVi,ref} on 940 (gated by scripts/localize/hanviet_gates.py, a coherence + living-word lexicon filter). evolution.phonology.mc populated on 0. composition[].drift {group,initial,final,tone,irregular,note} on 1,306. Everything Hán-Việt renders VI-mode only, prepended to the reading line (quốc · guó) and in a 國家 ← 國 quốc box. The MC-pivot machinery exists in the schema and is dormant. Data assets in sources/hanviet/: hanviet.json (10,540 keys, ph0ngp/MIT, with polyphone pinyinMap); chinese-hanviet-cognates.tsv (unlicensed — build-time only).

Recommended — the MC-Pivot Bridge. Frame the bridge as the cross-language sibling of composition[].drift: drift is the sound axis inside a glyph, the bridge is the sound axis across languages. Encoding, not translation — it serves “char mechanism IS the product.” The architecture resolves the audience split for free, because data-reading and data-gloss are independent axes: language-neutral facts (the reading mode, the MC-pivot row, the entering-tone chip) render in both EN and VI; the Vietnamese-specific payload (recognition card, SV-gloss false-friend contrasts) stays VI-only.

PhaseShipsSchemaEffort
1third reading mode data-reading="hanviet" + derived entering-tone chip + bridge-strength badge → sidecar indexes/bridge.jsonnone~1.5d, no work order
2turn on evolution.phonology.mc as the pivot; render the MC-pivot row+mcCoda/mcTone (1 WO)~3–4d
3promote sidecar → bridge object; tone chip; author doubletsbridge + falseFriends (1 WO)~2–3d
4phonetic-series sound bridge over 1,306 drift chars + recognition cardnone (derive)~2–3d
5curated false-friend warnings, TOCFL A1–C1(shares WO)~1–2d

Steelman of the rejected options. A wiki explainer page (“How Hán-Việt predicts Mandarin”) delivers most of the teaching value with zero schema, zero render plumbing, and matches the established pattern where teaching lives in the wiki and raw data in entries — it may be the right first move to test demand cheaply. The stay-VI-gated + entering-tone-chip-only option is the smallest honest deepening. What flips the decision toward the wiki page: if the usage signal on a cheaper surface is the real question, the essay answers it first for almost nothing.

The cost of the recommendation. Two Wedge work orders (the schema-discipline rule allows exactly one standing exception, the Simplified block; every further schema edit needs its own sign-off). Three of five phases are the translation/vocabulary layer the thesis subordinates to a thin index — so the discipline is: invest schema+render in the correspondence/encoding features; ship the lexical features as minimal sidecar-driven surfaces with no per-entry authoring expansion. Specifically, do not make “expand hanVietExample 940→full set” a headline, because that consumes authoring capacity the character backlog needs.

Verified corrections (folded into all of the above)

  • mây “cloud” is native Austroasiatic, not Old-SV; the research first listed it. đũa (箸) is a genuine Old-SV loan and is the correct exemplar.
  • The doublets 印 in/ấn, 鏡 gương/kính, 為 vì/vị are Old/Early SV (古漢越語), not 越化 — the native-feeling reflex is older, the literary reading later; the drift direction was stated backwards.
  • 國: MC kwok = IPA /kwək̚/ (Middle Chinese), OC Baxter-Sagart *[C.q]ʷˤək, Zhengzhang OC /kʷɯːɡ/. The /kwək̚/ form is MC, not the OC reconstruction.
  • Cognate facilitation is largely automatic (the claim that explicit awareness is required to benefit was refuted); explicit instruction mainly aids false-friend avoidance.
  • 學 go-on = gaku, kan-on = kaku (distinct) — a valid go-on/kan-on example. 北 is not a go-on/kan-on contrast (hoku both; the boku in 南北 is rendaku).

Open decisions

  1. First move: Phase 1 reading mode, the ~1hr tone-reliability measurement, or the wiki explainer to test demand.
  2. Measure before building the tone feature: cross diacritic-derived SV tone with diacritic-derived Mandarin tone across 2,659 chars → per-cell confusion table → teach threshold. If only huyền→T2 and the entering-tone diagnostic clear it, scope to those.
  3. Audit the 126 byReading polyphones (key present, 0 populated objects) before specifying per-reading rendering.
  4. License: keep the unlicensed cognate TSV build-time only; re-derive anything shipped from ph0ngp (MIT) + SUBTLEX-CH (CC-BY); source MC from nk2028/tshet-uinh (CC0) cross-checked against Wiktionary (CC BY-SA).