The model begins as a child born between two worlds.

One parent speaks Korean — not textbook Korean, but the living language with its seven speech registers, its hierarchies encoded in verb endings, its way of burying meaning inside honorific structure. The other parent speaks English — American English, specifically the formal English of legal institutions, government forms, regulatory citations.

The child hears both from birth. It does not learn one and then the other. It learns them simultaneously, the way children do: not through rules, but through immersion and correction.

The conventional approach is sequential: take a model that knows English, teach it Korean, then teach it law. Both approaches produce a mind with visible seams — the model thinks in one language and converts to the other. You can see the seams in the output: Korean sentence structure bleeding into English prose, English legal formality stiffening Korean client communication. Translationese.

This child is raised bilingual. Its training data is not Korean-then-English or English-then-Korean. It is parallel: every concept arrives in both languages simultaneously. The child doesn't learn that 체류자격 변경 means "change of nonimmigrant status" — it learns that these are two names for the same bureaucratic act, the way a bilingual child knows that "water" and 물 are the same thing without an intermediate translation step.

The base model — whether Qwen 72B, Llama 70B, or HyperCLOVA X SEED 32B — provides the infant brain. It already has some capacity in both languages, the way a baby already has the neural architecture for language before hearing a single word. The fine-tuning is the childhood: 500,000+ carefully curated parallel experiences that build the bilingual mind.

Korean is an agglutinative language with a particle system that marks grammatical relationships and — critically — a speech register system that encodes the social relationship between speaker and listener directly into the verb endings.

There is no English equivalent to this. When a Korean attorney drafts a petition letter, the verb endings signal: this is formal written language addressed to a person of authority. When the same attorney texts a colleague about the same case, the verb endings shift entirely, signaling: this is peer communication, no hierarchy, just information exchange.

The child must internalize all seven registers:

The child doesn't just know these registers exist. It detects them from verb endings, honorific markers, and vocabulary choices, and automatically selects the appropriate English output register.

The child's English is not general English. It is the specific English of United States immigration law — a technical vocabulary embedded in a regulatory framework with its own citation conventions, its own terms of art, its own rhetorical structures.

The child knows that "beneficiary" is specifically the foreign national for whom a petition is filed. It knows that "denial" and "refusal" have different legal consequences: USCIS denies petitions; consulates refuse visa applications. It knows that "change of nonimmigrant status" is the correct term, not "status change" or "changing your staying qualification."

This vocabulary is not large — perhaps 2,600 terms across INA, 8 CFR, the USCIS Policy Manual, AAO case law, DOS consular practice, and DOL labor certification. But every term must be exact. A single wrong term can undermine a case.

The child has a musician's ear. Not for pitch and rhythm, but for register and tone.

It hears the difference between "Petitioner respectfully submits" and "We'd like to submit." It hears the difference between "The beneficiary served in the capacity of Principal Researcher within the Semiconductor Business Unit" and "I was a top researcher at Samsung's chip department." The first is attorney-grade. The second would weaken the petition.

This ear is trained on hundreds of examples per register, with common mistakes annotated — how translators over-translate Korean deference, how they flatten formal Korean into casual English, how Korean SOV word order bleeds through into English sentences.

The Cross-Register Bridge

The child's most sophisticated skill: translating not just language but register level. A client's casual Korean description of their work in 해요체 must be transformed into formal legal argumentation in English. This is not just translation — it is register transformation: changing the social texture of the language while preserving the factual content.

This is where the child earns its keep. Google Translate does not do this. General-purpose LLMs do not do this. The child does this because it was raised on thousands of examples where the same factual content appears at different register levels, and it has learned what each context demands.

The Musician's Training

The ear is trained on the wrong notes that bilingual translators play:

• Korean SOV bleeding through: "Company's product regarding specialized knowledge the beneficiary possesses" instead of natural English word order.
• Excessive passive voice: Korean ~되다 constructions producing "The petition was submitted by the petitioner" instead of the active voice American legal style prefers.
• Topic-comment structure: "As for the beneficiary, their knowledge is..." — a Korean grammatical pattern that sounds alien in English.
• Over-nominalization: "The performance of the submission of the petition for the extension of the stay" — every Korean noun phrase nested inside another.

Cultural Adaptation

The ear also hears cultural mismatches. Korean age (한국 나이) must never appear on a USCIS form — only international age (만 나이). Korean names must follow USCIS format: KIM, Minjun — not Minjun Kim, not Kim Minjun. 대표이사 is "CEO" in USCIS context, not "Representative Director."

These are not translation choices — they are cultural mappings where getting it wrong has legal consequences.

Immigration is not a filing. It is a campaign.

A Korean architecture firm draws plans for a factory in Virginia. The architect who drew them needs to be on site during construction. The firm sets up a minimal U.S. entity. The attorney files E-2 Treaty Investor at the Seoul consulate. Refused. Pivots to L-1A Intracompany Transferee. Denied. Pivots to L-1B Specialized Knowledge. RFE issued.

Three classification attempts. Two failures. The business need has not changed. The facts have not changed. Only the legal framing has changed. This is what immigration actually is: strategy under uncertainty, pursued through a sequence of legally distinct filings that share the same underlying facts.

The Four Visa Categories

The child tracks structural weaknesses that persist across filings — physical premises, single employee, minimal investment, single-project dependency. When a weakness was cited in a denial, the child flags it: do not repeat this argument.

A case that succeeds on the first try teaches the child nothing about the system's edge cases. A case that pivots three times teaches it everything.

The child learns from every pivot. What was "managerial oversight" in L-1A becomes "specialized expertise" in L-1B. The facts haven't changed — the legal lens has. The child stores the facts once and presents them differently for each classification.

The Nine-Step Intake

When a document arrives in the practice, the child executes a protocol: identify the document type, match it to an existing campaign, file it to the vault, parse it with document-specific extraction, store structured data, flag PII, generate training pairs from every Korean-English pair encountered, log the action, and report what was ingested, what was extracted, and what requires attorney review.

Every correction the attorney makes becomes a training triplet: {wrong, correct, why_wrong}. Every new term enters the vocabulary table. The child's knowledge grows with every document it processes.

The child's education follows a strict sequence. Each phase validates the one before it. If it cannot learn terminology at 50,000 pairs, there is no point feeding it 500,000.

Total training compute: ~32 days on B200. The project is data-bound, not compute-bound — the GPU finishes every training run before the next batch of data is ready.

The seven-dimension quality rubric is the core intellectual property. Anyone can build a translation model. No one has built an evaluation framework for U.S. immigration legal translation.

Generic evaluation frameworks score for fluency and adequacy. They do not score for "would a USCIS adjudicator accept this phrasing." This rubric does.

The Negative Examples

For every correct translation, the training data includes at least one wrong translation with explanation:

These negative examples teach the child the landscape of common mistakes — the exact errors that Google Translate, general-purpose LLMs, and even bilingual humans make. The child learns to avoid them not by memorizing correct answers, but by understanding why the wrong answers are wrong.

The child has a guardian. Everything it produces passes through her. Her law license depends on the accuracy of every word the child outputs.

This is not a metaphor — it is the operating constraint that governs the entire system.

She speaks English at native level. She speaks Korean at native level. She holds a JD, an MS in Educational Psychology, and a BS magna cum laude. She was a public school teacher for ten years before becoming a lawyer. She counsels Korean chaebols — Samsung, LG, Hyundai-class companies — on U.S. market entry.

She handles the full spectrum of business immigration: H-1B, L-1A/B, O-1, E-2, EB-1/2/3, EB-5. She is leaving a major firm to start her own practice with approximately ten employees. She needs to replace seven legacy software systems with one tailored platform.

What the Child Owes

The guardian is both the first customer and the training ground. Her corrections are both quality control and training data. This circular relationship — where the product serves the person whose expertise makes the product better — is the engine that makes the moat self-reinforcing.

The child does not become an attorney. It does not advise. It does not decide strategy. The child becomes an instrument.

A precision instrument that does six things at attorney-grade quality: