Horizontal bands are strata; vertical columns are drill cores (case studies). Hover a core to light up the strata it crosses and where it flows out; hover a stratum to see which cores pass through. Click a core to enter its full record. — desktop hover; lists on touch.

02 · Circulation — the whole machine on one sheet · every box navigates 순환 전도

World — outside

↓ what flows in — erratics collected · inquiries received

Survey

the study wheel

new zones P · deepening zones S · erratic dissection

Production

the work wheel

Fields F1–F·X × Strata L0–L4 (↑ territory)

↓ judgment and collection produce raw stones

Field Notes

raw stones

unpolished specimens · promotion queue

↓ promotion — a stone becomes a question

Cores

measurement

scorecards · error taxonomy · intervention · re-measurement

⇄

破律

PAYUL — the shelf

shelf-bound questions ↔ material (1,293)

↓ what a core yields

↓ outflow — GitHub · LinkedIn · RSS · Threads · Medium

World — what flows out returns as inquiries (a closed loop)

Record = the cross-section of this entire loop ↗

03 · Gauge — watching that both wheels keep turning 계기 — 두 바퀴

Production — the wheel work turns

07-24Vendor review → 1 specimen into Field Notes (dummy)

ongoingCore #001 — error-taxonomy stage

Survey — the wheel study turns

08-planPCT Guide reading → P1 log (dummy)

weeklyS1 Economy — 1 close read → 2 terms indexed (dummy)

07-251 erratic dissected → to be cited in #001 (dummy)

Two-wheel principle — a mine that only produces is depleted; a mine that only surveys starves. The lab moves only when both wheels turn. When one stops, this panel goes empty — and shows it. RHYTHM: weekly note (vault · private) → this ledger · calendar = ICS subscription — personal schedule never enters this public site

04 · Outflow — only finished cores leave the lab 반출

GitHub

Core Archive

Where drill cores are stored: data, rubrics, sample prompts. English README, CC BY 4.0. Only the public layer leaves — you can publish the recipe; the tongue can't be copied.

PUBLIC: error taxonomy · method · before/after / PRIVATE: full rubric · production prompts

↗ github.com/──── · handle TBD

Showcase

Where core summaries and before/afters go on display, stacked on four anchors (Centific · Acclaro · Alconost · Snowflake).

Rule: no end-client names for ongoing contracts — "via [LSP]"

↗ linkedin.com/in/hyoroarkkim

ProZ

Gauge

The reputation dial: 5.0 · 13 reviews. Independent of core flow — wired straight to Record.

→ RECORD · client-name rule applies on WIWO too

↗ proz.com/feedback-card/2792527

RSS

Pipe

A quiet conduit straight into the lab — no algorithm, no metrics, no presence required. Cores and field notes flow as published. The quiet producer's default plumbing.

FEED: cores + field notes · newsletter re-judged after 3 published cores

↗ /feed.xml · to be built

Threads

Specimen Window 표본 창

A Korean-language window where short specimens flow. The site is always the canonical home — only what's already in Field Notes gets poured out. No separate writing = no presence pressure.

FEED: field-note excerpts · re-judged & adopted 2026-07 · same naming rule

↗ threads.net/@──── · handle TBD

Medium · dev.to

Outcrop 노두

The same stratum surfacing elsewhere: English core write-ups republished with a canonical link back here. dev.to for developer-facing cores only.

Rule: republish +1 week after site · canonical required · no comment/follow upkeep · re-judged & adopted 2026-07

↗ medium.com/@──── · handle TBD

Verdicts — rejected: X·Bluesky (demand presence and reaction speed) · YouTube (no-talks principle) · Brunch (no canonical support — the authenticity problem can't be solved). Re-judged & adopted: Threads (restricted to field-note excerpts) · Medium/dev.to (canonical republishing = outcrop — the very reason for rejection dissolved by canonical; by the same yardstick, Brunch stays rejected). Reddit·HN: one-off outflow events at tool launches only. Rejections and reversals are both verdicts, so both are recorded.

05 · Inbound — the door where work comes in (mirror of Outflow) 반입 — 의뢰

Inquiries

EN→KO — corporate communications · legal & compliance · IT & cloud · AI output evaluation (LQA · MTPE). Based in Seoul (GMT+9): your evening is my working night — overnight turnaround into your morning is the default rhythm.

What to send: source language and field · volume (word count) · deadline · a sample paragraph if you have one. Rates follow the document and scope — quotes within one business day.

✉ [email protected]

Test pieces welcome · vendor-form ready · NDA-disciplined — the same rules I keep for every ongoing contract

Specimens 표본 제보

Found a broken honorific, a snapped particle, a patch of translationese in the wild? Donate it to the lab — one sentence and where you saw it is enough. Accepted specimens are laid into Field Notes; credit optional (anonymous welcome).

✉ Send a specimen

Objections are specimens too — disagree with an Index verdict? Same door.

06 · Atlas — the cards are the table of contents 상세 지도

07 · Triangulation — three stations, one survey 삼각망

This lab does not stand alone. JEOKRYUL and PAYUL are the two known stations — sister grounds holding the same 律 with opposite verbs — and ATU is the new point: a work on PAYUL's shelf, an instrument and material (#004) in this lab. As in surveying, a new point gets its coordinates only when sighted from both stations at once. All three sites share this one chart; only "you are here" changes.

↗ 破律 PAYUL · URL TBD ↗ ATU · URL TBD

Cores — the drill archive

Core box laid in the tray in drilling order코어 박스 — 시추 아카이브

Each core is one question. The bar on the left shows the strata it crosses — darker is deeper. When a core finishes (DRILLING → LOGGED), it leaves as a pair: a GitHub repository and a LinkedIn write-up.

#001 · DRILLING · field: all · est. 2026-08

Honorific collapse — register consistency in LLM Korean

Where and how often the polite register (해요체) breaks down in long generation, measured on a scorecard; prompt intervention compared before and after. DeepL included among the models.

OUTFLOW: repo — pending post — pending

#002 · PLANNED · field: F1 IT

Particles × variables — the fate of {name}을(를)

A map of how object/subject particles break when variables are injected into UI strings — designed to drill all the way to L4, where the core becomes the particle checker's spec.

OUTFLOW: not started

#003 · PLANNED · field: all

A map of translationese — where English leaves its shadow

Passives, noun-phrase pileups, lingering pronouns — the translationese humans and machines share, typed and turned into candidate detection rules.

OUTFLOW: not started

#004 · PLANNED · PAYUL crossover

The 1,293 corpus — LLM Korean under a closed vocabulary書架 ↗

PAYUL's material as this lab's experimental condition: constrain generation to ATU's closed corpus of 1,293 words and measure what happens to honorifics, particles, translationese. The core that proves the two sites are one body.

OUTFLOW: not started · material source: PAYUL — Across the Universe

← back to the core box

Core #001 · Drilling · L3 → L0

Honorific collapse measuring register consistency in LLM Korean존댓말 붕괴 — LLM 한국어의 존대 일관성 측정

The longer the document, the more the machine's polite register wavers — formal endings slip in, plain forms leak out. Where, why, how often — and how much one line of prompt can hold it.

Pipeline

problemscorecardcollection error taxonomyinterventionre-measuregeneralize

Specimen (dummy) — Korean is the material; glosses follow

Before · raw model output — red-pen markup

계정을 만들려면 이메일을 입력하세요. 다음으로, 비밀번호를 설정합니다. ~~완료되면 확인 버튼을 눌러.~~ 완료되면 확인 버튼을 누르세요.

Gloss: three UI sentences. The first two hold the polite -요 register; the third drops to plain speech (눌러 "press-PLAIN") — register collapse at sentence three. The red pen restores -세요.

After · with prompt intervention

계정을 만들려면 이메일을 입력하세요. 다음으로, 비밀번호를 설정하세요. 완료되면 확인 버튼을 누르세요.

Gloss: sentence-final endings now consistent. Residual errors still observed inside conditional clauses — under measurement.

Error taxonomy — interim tally (dummy figures · 120-sentence sample)

Type	Description	Freq.	Representative specimen
Formal intrusion	-십시오/-습니다 endings leak into a -요 document	31%	"…입력하세요. 다음을 확인하십시오."
Plain-form drop	endings shed in lists and short sentences	18%	"버튼을 눌러." · "설정 완료."
Quotation loss	honorific features evaporate inside reported speech	11%	"관리자가 승인했다고 표시됩니다"
Over-elevation	casual brand voice escalated to archaic deference	7%	"곧 만나 뵙겠사옵니다"

Interim reading: collapse is not random — it clusters on three terrains: sentence length, list structure, and reported speech. The intervention stage will name these terrains explicitly in the prompt and re-measure. (Figures are illustrative.)

Method — public / private layers

Layer	Contents	Outflow
PUBLIC	error classification scheme, measurement method, before/after specimens, summary figures	GitHub · CC BY 4.0
PRIVATE	full scoring rubric, production prompt text	withheld — publish the recipe; the tongue can't be copied

Shelf-bound question — what crosses to PAYUL when the drilling ends

Why does the machine forget deference — and when do humans break it on purpose? Where measurement ends, an essay begins.書架 ↗

Field Notes — the collection log

Raw stones specimens before they become questions야장 — 원석

Specimens picked up in daily practice, logged with only a date and a stratum tag. Not polishing them is the rule — the format that keeps perfectionism from delaying the record. Some stones get promoted to cores (#). Korean text below is the material itself.

2026-07-26
L2 F3

Weather metaphors in earnings language — headwinds/tailwinds. Has the literal 역풍·순풍 actually settled into Korean IR practice? Needs a check against three filings. A metaphor's citizenship is granted by usage, not by the dictionary.

2026-07-25
L1 F3

"Doom loop" in The Economist — Korean financial press already favors the transliteration 둠루프. The trade of importing a metaphor untranslated: the concept survives, but a metaphor's body heat dies at the border. Same vein as the headwinds specimen (07-26) — the press isn't current affairs here; it's an outcrop of the L1 stratum. (dummy)

2026-07-24
L0 F2

Vendor review — every "please note that" rendered uniformly as "~라는 점에 유의하시기 바랍니다". Not wrong, but the whole document turns into an official gazette. An intensity scale by context is needed. → #003 candidate

2026-07-21
L0 L3 F1

MTPE evaluation specimen — in a three-item list, two items in polite -요, one in formal -습니다. Same screen, same prompt. The collapse isn't random; it looks correlated with sentence length. → promoted to #001

2026-07-19
L0 F1

Toast messages — 저장되었습니다 (saved-PASSIVE) vs 저장했습니다 (saved-ACTIVE). The passive doubling as humility in product voice. Is erasing the agent politeness or evasion? A question a brand-voice guide has to answer.

2026-07-17
L1 F1

"the cloud" vs "a cloud environment" — the reference range that articles split, Korean catches with particles and word order. Where the articles book (PAYUL) meets this lab. A crossover specimen. 書架 ↗

2026-07-14
L1 F2

Contractual "may" has three faces — permission, possibility, discretion. Flatten them all into "~할 수 있다" and a discretion clause becomes a capability clause. The twin of "shall" (already in the Index).

2026-07-11
L2 all

What "keep the tone" really asks for in transcreation — not preserving the source rhythm but replacing it with the target reader's expected rhythm. One paragraph explaining this distinction in client language is overdue.

2026-07-08
L0 all

The plural marker 들 in excess — the shadow of English -s. Korean lets context carry number, yet translations insist on 사용자들은. Specimen no. 1 for the translationese map. #003 translationese map · not yet drilled

2026-07-03
L2 F·X

"Practice" in an exhibition foreword — 작업 (one's work) or 실천 (praxis)? Some words can't be judged without the artist in the room. Deferral is also a verdict — in conceptual-art context the two renderings split the artist's self-definition.

Erratics — 漂石

The dissection table stones carried in from elsewhere, laid down with provenance표석 해부대

Geology calls a boulder the glacier carried in from another stratum an erratic — and its origin, provenance. Insights swept in from LinkedIn, Threads and the web are laid here: collect instantly and raw, dissect later in batches. Excerpts stay under two sentences; the body is always my dissection. Dissection triggers action, and the journey of execution stacks up on the card's own timeline (TRAIL). Four life stages — COLLECTED → DISSECTED → IN ACTION → SEDIMENTED (internalized): a foreign stone weathering until it becomes my own stratum is one erratic's complete life. The single sediment bed that replaced my scattered private notes.

WEB · The AI Journal — "Best AI Translation Tools in 2026" (Erika Balla) · collected 2026-07-26 · source ↗IN ACTION

Paraphrase: a 2026 roundup of five enterprise AI translation tools — Google Cloud Translation v3, Amazon Translate, Lara Translate, DeepL API, Azure Translator — with DeepL noted for European-language quality, glossaries, formality controls, and a clean API.

Dissection Genre first: the vendor-roundup format scores language count, API, speed, ecosystem fit — and has no row for register consistency or terminology drift under length, the axes where Korean actually breaks. The absent rows are my market. And one empirical hook: what DeepL's "formality controls" actually do to the Korean honorific system is not a marketing sentence — it's a thing to measure.

ACTION PLAN — what this dissection triggers

□ 1Get a DeepL API free-tier key + check official docs for the formality parameter's KO coverage → Bench BUILD LOG #005

□ 2Batch-run 20 honorific specimens — measure register retention per formality setting → #001 comparison-model axis

□ 3One hour observing Lara Translate's glossary behavior — judge whether it enters the Index watchlist → INDEX candidate

TRAIL — the timeline where execution stacks up

07-26Collected + dissected — three actions issued

07-27 planAPI key + docs check → logged as Bench #005

07-29 plan20-sentence batch → specimen table → into #001 §models compared

laterSEDIMENTED verdict — if DeepL becomes the standing comparison rig in my pipeline, this stone has become stratum

JOINTS#001 honorifics Bench Instruments

THREADS · a translator (pseudonym · dummy) · collected 2026-07-25 · source ↗CITED → #001

"존댓말 일관성은 프롬프트로 안 잡혀요. 시스템 프롬프트에 세 번을 써도 목록만 나오면 무너짐." (excerpt — "Honorific consistency can't be fixed by prompting; write it three times and it still breaks the moment a list appears.")

Dissection A field intuition that lands exactly on my interim tally — list-structure collapse (18%) does resist prompt repetition. But my hypothesis sharpens it: not "can't be fixed by prompting" but "can't be fixed without naming the terrain." Queued as a counter-example candidate for #001's intervention stage.

JOINTS#001 honorifics note 07-21 L0 bedrock

LINKEDIN · a localization lead (dummy) · collected 2026-07-23 · source ↗DISSECTED

Paraphrase: the practice of pricing MTPE as a percentage of translation shifts the quality risk of variable MT output onto the linguist.

Dissection Looks like a rate debate; it's actually a risk-allocation problem — when output-quality variance is high and the rate is fixed, the linguist absorbs the variance cost. A ready-made frame for my own floor-price principle ("their range first"). Counter-point worth making: whoever can measure output quality in advance (= this lab's scorecards) holds the negotiating position.

JOINTScore methodology Record — rate discipline F1

WEB · industry article (dummy) · collected 2026-07-20 · source ↗DISSECTED

Paraphrase: an LSP survey finding that terminology consistency has become harder to manage after LLM adoption than it was in the TM era.

Dissection A good erratic because it runs against intuition — TM enforces exact matches; LLMs produce plausible variation. The enemy of consistency is no longer error but fluent variance. Which is precisely why the Index records the verdict behind each term, not just the rendering.

JOINTSINDEX L3 machine

PREPRINT · abstract on honorific modeling (dummy) · collected 2026-07-18 · source ↗COLLECTED — awaiting dissection

Paraphrase: an experiment reporting improved generation consistency when Korean honorific features are handled as a separate token layer.

Dissection — pending Collected only. To be handled in the weekly dissection batch — laying it down raw was today's whole job.

JOINTS#001 candidate reference candidate

Correlation — 對比圖

The joint web drawing lines between drill logs대비 — 접점 그물

Geologists lay columnar sections side by side and draw correlation lines between matching strata. Here the content itself is the columns: erratics shake cores, notes get promoted, cores yield an index and tools — and only those real events become lines. Hover a node to light its web. This chart is the public rendering of my Obsidian graph — the inside window and the display window render the same links (vault = site repository).

Erratics · from outside

THREADS · CITED"honorifics won't hold by prompt"

LINKEDINMTPE rates = risk allocation

WEBLLM terminology — fluent variance

PREPRINThonorific token layer

Field Notes · raw stones

07-24 · F2uniform "please note that"

07-21 · F1register collapsing in lists

07-17 · F1the cloud / a cloud — articles

07-08 · allplural 들 in excess

07-03 · F·X"practice" — deferral is a verdict

Cores · measurement

#001 · DRILLINGhonorific collapse

#002 · PLANNEDparticles × variables

#003 · PLANNEDtranslationese map

#004 · CROSSOVERthe 1,293 corpus

Yield · assets

INDEXprovision — one word, three faces

INDEXtranslationese rule candidates

INSTRUMENThonorific-collapse detector

INSTRUMENTparticle checker

破律 · SHELFarticles book · ATU — the mirror room

Same star topology as 破律 v9 — nodes in one column never connect directly; only promotion, citation and yield become lines. A node without lines ("practice") isn't unfinished: it's a stratum not yet correlated. In the inside window (Obsidian) this web draws itself — vault = site repo.

Index

Term strata not just the rendering — why this rendering색인 — 판정의 지층

A glossary that buries the verdict alongside the term. Searchable assets, and the raw ore for the L4 detectors. Korean renderings are the material — glosses in the memo column.

Term	Rendering (KO)	Verdict memo	Str.	Field
observability	관측 가능성	Must be kept distinct from 가시성 (visibility) — mixing them in monitoring docs collapses the concept. Translate-verdict.	L2	F1
hydration	하이드레이션	In frontend context the literal 수화 (hydration-as-chemistry) misleads — transliterate-verdict; source term glossed at first mention.	L2	F1
ephemeral	임시 / 휘발성	Storage context: 임시 (temporary); container context: 휘발성 (volatile). Refusing a single fixed rendering — context-split verdict.	L2	F1
you	(생략) / 사용자	A literal second person collides with the honorific system in Korean UI — subject omission is the default; 사용자 (the user) when needed. An L0 verdict.	L0	F1
deprecated	지원 중단 예정	사용되지 않음 ("not used") misreads as current state — a tense-and-aspect (L1) problem deciding a terminology choice.	L1	F1
shall	~하여야 한다	Contractual shall is obligation — render it as plain future ("~할 것이다") and the legal force changes. Not a modal; a clause.	L1	F2
guidance	가이던스 / 지침history: 2026-05 single rendering → 2026-07 split by field (IR convention verified)	IR context (earnings outlook): transliterated 가이던스 is market convention; regulatory context: 지침. A case of the field deciding the rendering.	L2	F3
provision	프로비저닝 / 조항 / 충당금history: 2026-06 IT-only entry → 2026-07 three-way verdict (origin of the field tag)	One word, three faces across three fields — IT transliterates, law says clause, accounting says reserve. Why the field tag exists.	L2	all
indemnify	면책하다	Never conflate with 배상하다 (to compensate) — the direction is opposite (protection from third-party claims). The densest mistranslation zone in contracts.	L1	F2
hereinafter	이하 "…"(이)라 한다	Replaced not word-for-word but with the Korean contract's definitional formula — structural translation; a literal rendering breaks contract grammar.	L1	F2
workload	워크로드	작업 부하 (work-load-as-strain) only in performance contexts — as a cloud resource unit, transliteration is the standard across major vendors' Korean docs.	L2	F1
run rate	런레이트	IR-convention transliteration — unpacking it as 연환산 fixes one definition and misleads. Parenthetical definition at first mention.	L2	F3
prior art	선행기술	Fixed statutory term — KIPO's solid-spaced 선행기술. Collected in survey zone P1 (dummy).	L1	P1

Disagree with a verdict? Send an objection. Objections are specimens — accepted ones update the memo, with credit (anonymous welcome).

Instruments

Judgment, tooled the terminus of the arc도구 — 판단의 도구화

Detection rules extracted from cores become tools. Time → judgment → assets → authority — this stratum is where the arc lands. The making-of lives on the Bench — friction, code, and musings.

Try it — the Ear Test · the human demo of the detectors

Before the rules catch it, try your ear. In each pair, pick the sentence where the machine left its trace. Korean is the material; every answer comes with a gloss. Scored instantly, nothing recorded.

1 / register consistency

The second option drops from polite -세요 to plain 눌러 mid-passage — the "plain-form drop" Core #001 is measuring (18% in the interim tally).

2 / particles × variables

The moment a final-consonant word ("문서.txt") lands in the variable, option B breaks ("문서.txt를") — exactly what the particle checker will catch. A shows the safe dual form.

3 / translationese

Plural markers and a possessive pronoun shadowing English -s and "their" — field note 07-08, the terrain of Core #003.

0 / 3 — pick away. Your ear is the first instrument.

SOON · from #001

Honorific-collapse detector

Scans a whole document for sentence-final register consistency — flags drops to plain speech and formal intrusions. Core #001's error taxonomy is the source of its rules.

IN: text/doc → OUT: register map + drop coordinates
RULE e.g.: flag -십시오 intrusion inside a -요 document (illustrative)

✉ Join the waitlist

SOON · from #002

Particle checker

Validates 을/를·이/가·과/와 alternation in variable-injected UI strings — replacing a repetitive QA chore with a rule.

IN: UI strings + variable schema → OUT: particle-branch report
RULE e.g.: propose final-consonant branching over "{name}을(를)" (illustrative)

✉ Join the waitlist

LIVE · PAYUL crossover

Across the Universe書架 ↗

An English-learning app that crosses a universe on a closed corpus of 1,293 words. This lab's experiment became that shelf's work — it lives in both webs at once.

CORPUS: 1,293 words · 36 units · 1,075 sentences — the material of Core #004

↗ to the app · URL TBD

Bench — bench notes

The bench log where the tools get made — friction, code, musings작업대 일지

A record of building the app and the detectors, and of what broke on the way. The form is fixed — problem → friction → code → fix → muse → joints. Code is first-class material here, and the last two slots are what keep this a lab rather than a dev blog: the moments when writing code turns into noticing language. (Copy and code below are dummy.)

2026-07-20ATUdebuggingBUILD LOG #004

Thirty minutes frozen at a red screen — or, error messages are front-loaded prose

Problem: Unit 7 loads to a blank deck and a console full of red.

Friction — honest version: the first thirty minutes weren't debugging. They were fear. I sat in front of the red text touching nothing, certain one wrong move would collapse everything — the old dread of the non-engineer in engineer country. Then I caught what I was actually afraid of. Not the code. A text I couldn't read. And sitting down in front of texts I can't yet read is what I've done for a living for twenty-five years. So I treated the error like a source text. Line by line. Close reading.

console — the stack trace, read as a paragraph

TypeError: Cannot read properties of undefined (reading 'ko') at renderCard (deck.js:41) ← the conclusion comes first at loadUnit (deck.js:18) ← everything below is evidence

Reading it, I laughed. This is front-loaded English prose — the inverted pyramid I've been editing my whole career. Topic sentence on top (what died), supporting detail descending. The trace had been unreadable not because I don't know code, but because I didn't know the genre's information structure.

deck.js — before / after

// before — asserting presence of something absent const card = corpus[unit].words[i]; render(card.ko); // .ko = the Korean gloss field // after — letting absence be absent if (!corpus[unit]) return; // 90% of the fear ended on this line

Fix: root cause was rendering ahead of the fetch — one await. The guard stays as a seatbelt. Unresolved edge: still no loading state between the blank screen and the error screen.

Muse I didn't write the first draft of this code. An AI did; I broke it, rebuilt it, and passed judgment. That used to feel like a confession. Today it clicked: I've spent twenty-five years taking responsibility for manuscripts I did not write — commissioning, shaping, deciding what stands. An editor owns the book without writing the sentences. Vibe-coding is commissioning; the authorship is in the judgment. And the error message is the most honest author I've ever worked with — it states what's wrong in the first line.

JOINTSL1 — information structure INDEX

2026-07-22ATUJS · Korean unicodeBUILD LOG #003

One final consonant decides the particle — hasBatchim

Problem: when words drop into sentence templates, the object particle 을/를 breaks. Friction: how to know a syllable's final consonant without a dictionary — until the references led to the arithmetic inside Korean unicode.

atu/josa.js — final-consonant test

// Hangul syllable = 0xAC00 + (initial×21 + vowel)×28 + final // final index 0 = no final consonant — the whole particle branch is one modulo function hasBatchim(word){ const last = word.charCodeAt(word.length - 1); if (last < 0xAC00 || last > 0xD7A3) return false; // outside Hangul (digits, Latin) return (last - 0xAC00) % 28 !== 0; }

Fix: templates now branch through josa(word,'을','를'). Words ending in Latin or digits ("파일.txt") fall outside the syllable block — left as an unresolved edge needing a rule table.

Muse Centuries of particle grammar collapsing into one modulo. The particle checker (L4, SOON) was already hiding inside ATU — tools aren't invented; they're discovered on the bench.

JOINTSparticle checker #002 note 07-19

2026-07-09DETECTORregex prototypeBUILD LOG #001

Where the rule leaks is where measurement begins — a register scanner

Problem: can honorific collapse be caught by rules alone? Friction: the sentence-level regex took minutes to write — and failed in exactly one place.

proto/honorific-scan.js — v0

const FORMAL = /(습니다|십시오)[.!?"']*$/; const POLITE = /(세|해|돼)?요[.!?"']*$/; // sentence-level scan — but a list item isn't a "sentence": // "설정 완료" (setup complete) has no ending at all. Here the rule leaks

Partial fix: a list-context flag — but "where does the list end" is not a rule. It's a judgment.

Muse The boundary line of rule-based detection lands exactly on Core #001's measurement site (list-structure collapse, 18%). Where code stops, measurement begins; where measurement ends, code resumes — the circulation between bench and core.

JOINTS#001 honorifics detector

Prospects — the survey

Survey zones where study lives before drilling탐사 구역

This page is the whole of the study wheel. Zones come in two kinds — new zones (P, dashed) survey territory not yet worked; deepening zones (S, solid) dig under the strata already worked. The form is identical: read a document, leave a one-line verdict, bury the terms in the Index. P-zones graduate into strata on the first paid job or first core; S-zones never need to — they already belong.

New zones — territory before drilling (P series · dashed = provisional)

ZONE P1 · SURVEYING · opens 2026-08

Patent · IP

The adjacent vein of the US legal track — EN↔KO patent translation carries real, high-rate demand in litigation, examination, and PCT national-phase entry. One block over from the legal strength.

References — what gets read: WIPO Pearl (multilingual patent terminology DB — the Index's external benchmark), the PCT Applicant's Guide, KIPRIS (Korean patent search — real claims in Korean), WIPO drafting guides. Every reading lands in the log below.

Graduation: one patent job or one patent core → joins L2 · the "patent" Index tag goes official · border turns solid.

Deepening zones — digging under working strata (S series · solid = active)

ZONE S1 · DEEPENING

Economy · Finance

Ground work under corporate-comms and IR translation. The Economist editorial history is the benchmark — close reads set beside the language of filings (DART · SEC).

FEEDS: Index "economy" tag · study twin of field F3

ZONE S2 · DEEPENING

Legal theory

The theory bed under compliance practice — clause structure, common-law basics. Letting theory explain the sentences practice keeps meeting.

FEEDS: Index "legal" tag · study twin of F2 · borders P1

ZONE S3 · DEEPENING

IT fundamentals

The concept bed under cloud and developer-doc translation — architecture papers, RFCs. From someone who moves the terms to someone who knows the concepts.

FEEDS: Index "IT" tag · study twin of F1

Reading log — one document, one entry, one-line verdict required (P/S alike)

placeholder
P1

PCT Applicant's Guide, international phase — WIPO. One-line verdict: the real translation demand sits not in filing itself but in national-phase entry — specification and claims at the designated-state deadline. (Format sample — replaced when actual reading begins.)

2026-07-27
S3

AWS Well-Architected Framework overview — verdict: the "pillar" metaphor system survives only halfway into the Korean docs. Metaphor loss becoming concept loss — a specimen. (dummy)

2026-07-25
S1

The Economist, one Leaders piece — verdict: Korean's landing spot for "soft landing", 연착륙, is already a dead metaphor. Translating a live metaphor into a dead one is itself a strategy; the real question is when it's allowed. (dummy)

Musings — thoughts before they're tidy are still records

placeholder
P1 L0

Korean in patent claims has no honorific problem — instead it has the information structure of subjectless, single-sentence monsters: legal Korean at its limit value. Due to meet the translationese map (#003). → core seed

placeholder
S2

"Including but not limited to" and the Korean contract's 등 — same job, read differently by courts. A memo comparing the binding force of illustrative enumeration, one precedent each side. (dummy)

Term collection

Terms collected while surveying go straight into the Index under a zone tag — study becoming asset through the same plumbing. Collection counts even before graduation.

Record

A 25-year cross-section nine years of localization over fifteen of publishing기록 — 25년 단면도

The geological CV of this lab. The top is recent; deeper is older — and the older strata carry the judgment of the layers above.

2016 – present · Localization & AI Evaluation

The localization stratum — 9 years

2026 –

Centific — Language Quality Lead, a major global cloud services account

2025 – 2026

Snowflake, Global Localization Team — client-side vendor review; MTPE evaluation and prompt-adjustment proposals inside the in-house AI translation platform

2024 –

Microsoft program via Acclaro (Azure · Copilot · Microsoft Learn) / legal & compliance via Andovar

2019 –

Alconost, 7+ years — JetBrains (ongoing), software & games localization

2019 – 2021

Flitto — Korean localization of Unreal Engine (Epic Games) & Unity developer-education content; AI language data

2017 – 2020

RWS (ex-SDL) — high-volume production across 25+ global accounts; the origin of the switching speed

FIELDS

F1 IT — Google Cloud · AWS · Autodesk · Hexagon / F2 Legal — sustained compliance programs / F3 Economy — The Economist (editorial & licensing) / F·X — European Commission · UNESCO · art criticism

2005 – 2019 · Publishing & Editorial

The editorial stratum — 15 years · where the judgment was trained

2020 – 2025

Gilbut — conceived and edited a translation-methodology title selected as a 2024 Sejong Book (a national distinction); negotiated the licensing deal with The Economist in person

2015

KPIPA delegation, Frankfurt Book Fair — presented "Reaching the Hybrid Reader" (ONIX · XML · discoverability)

2009 – 2015

SISA Publishing — 40+ English-education titles; licensing with Taiwan and Japan

2005 – 2006

Jungle Communications — 10+ design-industry programs incl. a Central Saint Martins collaboration — a career that began in design

Credentials · bedrock

Translation & Interpretation — Ewha GSTI, Seoul

Yonsei University

TOOLS

Phrase · Trados · memoQ · XTM · Crowdin · Smartcat / QA: Xbench · Phrase QA / Wrike

GAUGE

ProZ 5.0 · 13 reviews — feedback card ↗

Inquiries & CV

To hire the owner of this cross-section — EN→KO, with field, volume, and deadline. A conventional CV of this page exists for forwarding and ATS systems.

✉ [email protected] ↓ CV (PDF) — August edition, pending

Public PDF carries email only (no phone) · role-targeted versions (IT/enterprise · games · AI data · corp-comms/legal) available on request

Disclosure rule — end-client names of ongoing contracts never appear on this site. Only concluded contracts are named (e.g., Snowflake, ended 2026-07); ongoing work reads "via [LSP]" or stays anonymous. Not a limitation — the trust structure of this profession.