Upstage has fine-tuned its Solar family of language models on two years of Korea Exchange disclosure templates and is piloting machine-generated Korean summaries atop linked DART filings, the company and two broker research vendors said Sept. 18. The summaries do not replace official filings; they appear as a secondary pane for terminal users once a human reviewer clicks approve.
Which filings entered the training set
The corpus emphasizes recurring forms—quarterly earnings, material event reports, treasury stock disclosures, and convertible bond notices—rather than one-off prospectuses heavy with legal boilerplate. Upstage stripped tables into structured cells before training so the model learns row semantics instead of memorizing pixel layouts from scanned PDFs.
Roughly 180,000 documents from 2024 and 2025 form the core, augmented with synthetic examples where interns masked company names to reduce memorization risk. Auditors from a participating brokerage said they rejected 4 percent of generated abstracts in the first week for missing a footnote reference.
How the KRX pilot is wired
Terminal partners pull the official PDF from the Financial Supervisory Service's DART repository, hash it, then request a summary only if the hash matches a cached filing. That prevents the model from hallucinating against draft uploads traders might circulate on chat apps.
Generated text must cite paragraph anchors—"Item 3.2" style pointers—so compliance officers can jump to the underlying passage. Korea Exchange staff stressed that the exchange logo and statutory wording remain untouched; summaries are labeled as vendor commentary.
Human review rules
Every summary touching earnings guidance, auditor changes, or litigation triggers mandatory review by a licensed analyst before publication inside the terminal. Batch jobs overnight handle routine treasury share cancellations with spot checks on ten percent of rows.
Upstage hosts the inference cluster in Seoul and logs reviewer IDs for each approval, a requirement brokers imposed after internal simulations showed unreviewed summaries could misstate dividend dates when tables spanned midnight UTC boundaries.
Accuracy metrics shared under NDA
Two vendors allowed reporters to see aggregate scores: 92 percent entity-level accuracy on issuer names and ticker symbols, 87 percent on numeric fields when rounded to Korean won billions, and higher error on complex hybrid bond covenants. Those numbers are pilot-only; neither Upstage nor KRX published them on public websites.
False negatives—missing a going-concern warning—are weighted heavier than verbose phrasing. Reviewers said the model already beats junior associates on speed for boilerplate filings but still stumbles when companies mix Korean and English titles in the same table.
Regulatory context
The Financial Supervisory Service has encouraged AI assistance for reading disclosures but warned in spring guidance that vendors cannot imply regulatory endorsement. Upstage's UI uses gray banners reminding users that only the DART PDF is legally binding.
Retail investor protection groups asked whether summaries could front-run official translations for foreign retail holders; brokers responded by gating English summaries behind the same delay window as Korean ones.
Competition from banks' in-house models
Major securities houses already run internal GPT-style tools on research notes, yet few shared models across competing terminals. Upstage bets a filing-specialized weights file—updated monthly with new form codes—outperforms general models that treat covenants like generic legalese.
Naver's HyperCLOVA and NC's VARCO offerings were evaluated by one pilot broker and reportedly trailed on table-heavy extracts, though those benchmarks were not released publicly.
Commercial terms
Licensing is per-seat for terminal users with a surcharge when reviewers disable human approval on low-risk forms. Upstage declined to share pricing; a research vendor said it expects to pass costs to institutional clients only, keeping retail terminals on manual abstracts for now.
If the pilot extends beyond December, KRX may open a public comment period on whether summary panes belong inside the official disclosure viewer or only on broker overlays—a distinction that affects liability when numbers disagree.
Upstage engineers said they are testing English abstracts for dual-listed issuers next, but only after Korean reviewers sign off on anchor citations, because ADR holders still sue over mistranslated guidance numbers.