Executive Summary

What happened to AI Hub in 2026 was four jobs of four different kinds: counting the data the state already holds (the census), rebuilding the repository itself (the system procurement), taking in data from outside (the call for submissions), and reopening data that had already been taken in and published (the re-verification contract). Each ran on its own notice and its own budget, and each was reported separately. Put them in one place and what this repository did over the year reads as a single line.

AI Hub is the repository that, since 2017, has collected and published AI training data built with Korean government money. It is operated by the National Information Society Agency (NIA) under the Ministry of Science and ICT. Article 15 of the AI Framework Act, in force since January 22, 2026, gave the repository a new legal standing, and its name is changing with it. The formal name is the Integrated Provision System for AI Training Data, shortened here to the integrated provision system. The phrase integrated provision framework, which appears in the April coverage and in the name of the re-verification contract, points at the same thing. Original wording is kept only inside quotations.

Table 1. The 2026 timeline around AI Hub. The orange rows are the three this article follows. Current as of November 8, 2026.

One thing snags when you read that timeline. The phrase "100 datasets" pointed at four different things in 2026. Read two of the articles side by side and they look like one program, but the money and the responsible party differ. So the first job is to separate them.

Table 2. The four things "100 datasets" pointed at in 2026. The fourth row is handled separately in our article on the terms of use attached to those 100 high-value public datasets.

April's KRW 6 billion and the August notice's KRW 60 million per dataset multiply out exactly, which makes them easy to read as one pot of money. No document says they are. The August notice carries no total budget, and April's KRW 6 billion is attached to 100 datasets to be chosen inside the public sector. This article stops short of adding them and records only this much: the same "KRW 60 million per dataset" condition is attached to all three intake windows — public, private and university.

In August 2026 the Board of Audit and Inspection flagged duplicate construction and verification gaps in public AI training data, and that finding sits behind this overhaul. The audit itself is covered in a separate article. This one looks at what came after it.

What the state puts up in the August 3 notice is not a price for data. What the notice lists is technical and financial support — quality work such as consulting, plus de-identification — capped at KRW 60 million (about $41,000) per dataset for up to 100 datasets. Press-release distribution and media promotion are on the list of support items too. A company receives not cash but hands to work on its own data. And in exchange for those hands, what the state receives is not ownership but openness.

The university-sector notice states that condition more plainly.

The whole bargain fits in one sentence. Agree to open the data and the state covers the cost of cleaning it; decline, and the support goes away. From 2017 through 2024, what the government mostly did was build new data. The Board of Audit's tally for that period puts newly built public datasets at 908, worth $1.13 billion (KRW 1.63 trillion). The windows that opened in 2026 are a different animal. The money is now attached to fixing data somebody already holds and moving it into a published state. What building costs and what cleaning costs are gathered into a four-row table in Section 5.

So what happens to data too sensitive to open? The notice points at two controlled spaces. One is the Safe Zone run by AI Hub, where the data never leaves and can only be used inside a designated cloud environment. Access goes through a review that the notice says takes up to two weeks, plus up to five days to set up the environment. The other is the Data Safety Zone established under the Data Industry Act. The two are separate schemes with different enabling laws and different operators, and the similar names get them mixed up often. Four new Data Safety Zones were designated in May 2025, bringing the count to 14 zones across 11 institutions; any change since then is beyond what this article verified. On the Data Safety Zone side, the K-ICT Big Data Center reopened on October 1, 2026 as the AI Data Safe Use Center, taking on the role of central hub.

Here is the list of what a participating company must hand in, copied from the private-sector notice. Read the list alone and it looks as though the state has pinned down the definition of AI-ready in the form of paperwork. But one more line sits underneath the list.

Count the lines and only two are compulsory: the data specification sheet and the sample data. The remaining five sit inside a single bundle called "related deliverables," trailed by an "etc.," and the whole bundle carries the caveat "submit only those materials that can be submitted." In other words, the quality management report is optional, not required. The construction report that records how the data was built, the labeling guide and the metadata definition document all sit in that same box.

The university side is simpler still. The submission method the notice gives is to fill in the data specification form (Annex 3) and send it by email, and no support amount appears in the body of the notice. That does not mean there is no money; it means the body of the notice does not say. The private window's mandatory paperwork is two documents; the university's is one form.

One common misreading is worth heading off here. This figure is not a picture of missing standards. If anything the standards run thick. On February 27, 2026, NIA published the AI Data Quality Management Guidelines at v4.0: Volume 1, AI Data Quality Management Guide v4.0; Volume 2, AI Data Construction Guide v4.0; Volume 3, Generative AI Data Quality Management Guide v2.5; and Volume 4, Emerging-Technology AI Data Quality Management Guide v1.0, which covers data for reasoning models. The university notice attaches Volume 1 directly as Annex 4. The documentation has reached four volumes, and yet collecting the deliverables those documents call for is handled at the door with "hand it in if you have it." This is where the present article parts ways with our earlier piece on the absence of standards.

Academia already has a specification for what should travel alongside a dataset. The item list proposed in Datasheets for Datasets (2021) is the most widely used. That list carries a section called Maintenance, which asks only about life after the data is built. Those questions, set item by item against the paperwork collected at Korea's 2026 intake windows, fall into two groups.

Table 3. The Maintenance questions from Gebru et al., Datasheets for Datasets (CACM, 2021), set against the paperwork collected by the two 2026 notices. "Not found" is not the same as "absent": it means we could not locate it in the body of either notice, and the attached form files were not opened for this article.

A small footnote. The body of that guidelines notice says "three volumes in total," while four are attached. The files were reposted on April 15, 2026.

On August 27, 2026, the Ministry of Science and ICT and NIA put training data from the Sovereign AI Foundation Model project onto AI Hub. The press release that day set out what had been done before release: NIA and the Telecommunications Technology Association (TTA) ran quality verification, personal-information and harmfulness checks were carried out, and data under licensing constraints was held back. Twenty-two datasets went public. The press release and the AI Hub listing count that release differently, a problem taken up separately in our earlier article on that opening. This section does not recount them.

Five days later, on September 1, four of the 22 were switched to private. The stated reasons were inadequate de-identification and inadequate data structure. Roughly two weeks after that, on September 17, NewsPost published an exclusive on what it found after downloading and opening one of the four. It was customer-service conversation speech data built by NC AI, roughly 3,000 dialogue pairs and more than 500 hours of audio. What the reporter confirmed inside it were the names, phone numbers and addresses of real people.

The most interesting part of this episode is that two organizations called the same data by two different names. The report carried an explanation from each side.

Table 4. The two sides' explanations as carried in the NewsPost report of September 17, 2026. The quoted wording is translated from the Korean as printed in that report.

The two explanations look like they contradict each other, and they do not. They are two names for the same thing. The scripts were invented, and the people who read those scripts into a microphone were hired workers. One side named the recording process and said "workers recorded it themselves"; the other meant that no original customer calls were used and said "synthetic data method." Both can be true. And real identifiers still ended up in the output. The lesson that survives here is not which organization misspoke. It is this: "synthetic" is a label, not a guarantee that personal information is absent.

That proposition is not confined to one incident. The academic literature records two routes by which a synthetic data generator leaks its source. The first is verbatim memorization. When tabular data is produced with a large language model, numeric strings from the training examples can be reproduced exactly, and an attack works backward from the generated table alone to determine whether a given row was in the original (PoPETs 2026). The second is overlap in structure and neighborhoods. Even with differential-privacy noise added, cluster centers and neighbor relations stand in for the dense regions of the original and membership leaks. How badly this happens varies sharply by generator: under the same conditions, TabDDPM raised membership-inference success by 13.5 percentage points while CTGAN raised it by 0.5.

The contract made public by the September 29 report is titled, in translation, "Personal Information Re-Verification and De-Identification for Already-Released Data of the Integrated Provision Framework and AI Hub." The budget is $207,000 (KRW 300 million) and the period runs roughly two months, from contract signature to December. The scope covers 150 released AI Hub datasets mixing text, images and video. The premise written into the notice matters: these 150 are data that already went through a first round of de-identification and was published. The job is to look again for personal or sensitive information still inside. What gets inspected is not only text such as names and contact details but faces and license plates captured in images and video.

The scope of work is written in four stages. First, samples are drawn per dataset and checked automatically and manually for anything left behind; datasets flagged there are then inspected in full; additional de-identification follows, tuned to the characteristics of each data type; and finally the adequacy of those measures is verified. As deliverables, the contract requires an individual results report for each of the 150 datasets, holding the detection results, before-and-after samples and the techniques applied. The schedule puts detection, de-identification and full inspection in November, and report writing plus the final dataset upload in December. A further condition states that no GPUs, cloud or servers will be provided to the contractor, so the inspection infrastructure is the contractor's own to arrange. The winning bidder and the award price were not verified for this article.

The number 150 only means something with a denominator attached. As of November 8, 2026, the AI Hub data listing displays 977 datasets in total. What is being reopened is 150 out of a repository that size, and this is not a re-screen of everything AI Hub holds. The denominator itself, though, is not fixed at one number. One source puts the same platform at 945 datasets as of the end of 2025, and April 2026 coverage wrote 903 across 14 domains. Registrations keep accumulating, so a different counting date gives a different number. That is why this article does not compute 150 as a percentage of the whole.

Nor is the basis on which those 150 were chosen verifiable within the published coverage. One more thing snags on scope. The contract title says "already-released data," while the scope wording the coverage reproduced reads "data that is subject to, or a candidate for, the integrated provision framework." Whether all 150 are already public or whether candidates yet to be uploaded are mixed in could not be settled, because the original notice was not opened. This article reads it as the title does — already-released data — while recording that the boundary does not resolve from the coverage alone.

Per dataset, this contract's budget is KRW 300 million divided by 150, or about KRW 2 million ($1,400). The private-sector support ceiling in the previous section was up to KRW 60 million ($41,000) per dataset. It is tempting to set the two numbers side by side, but their natures have to be written down first. One is a ceiling that covers full remediation; the other is a budget-derived average for sample-based re-verification. They buy different things and involve different amounts of work. Table 5 gathers four such figures, with the nature and the denominator written onto each.

The second row needs its denominator flagged before you read it. It divides KRW 15 billion by 29 datasets — and 29 is the press release's count, while the 22 used in Section 4 is the AI Hub listing's count. Since that is the same release counted two ways, this article does not recount it and simply uses the press release's denominator. Switch the denominator to 22 and the per-dataset figure becomes about KRW 680 million ($469,000). Either way, the row's nature as the cost of building something new does not change.

Table 5. Four figures attached to one AI training dataset in Korea in the same year. The four rows differ in fiscal year, budget line and the nature of the work, so they are not added together. Nor are they divided into one another. The top two and the bottom two were priced for different jobs from the start. USD conversions use roughly KRW 1,450 to the dollar.

The same four figures also go into a chart. Top and bottom are close to three orders of magnitude apart, so on a linear axis the bottom two rows collapse into a single hairline. The horizontal axis is therefore logarithmic, and the price of that choice is that bar lengths are no longer quantities you can compare. Read the printed numbers and the nature label beside them rather than the lengths.

Finding personal information after publication is not a Korean phenomenon. Large English-language training corpora went through the same thing. A study that audited C4 — 365 million documents, 173 billion tokens — and The Pile — 22 bundled datasets, 383 billion tokens — after they had been distributed turned up personal information at the scale of millions of instances in both. Among the personal-information instances detected in C4, non-name types account for under 2.7%, and about 2.5% of those had another type of personal information sitting nearby. Both percentages take detected instances as the denominator, not documents; this is the spot where secondary citations routinely blur. Information clustered that way raises the risk of re-identification by linkage, and The Pile fared worse on it. A follow-up study in 2025 summed it up in one line: no automated cleaning removes all personal information.

So this re-verification contract is less a cleanup after an accident than the kind of work anyone running an open data repository ends up doing. What is different is that the Korean side wrote it into a procurement notice and put it out to tender.

On April 21, 2027, the premise changes. That is the day the amended AI Framework Act takes effect and connecting AI training data built with government money to the integrated provision system becomes mandatory. Every window seen so far has been voluntary: those who wanted to contribute did, and agreeing to open the data bought the cleaning work. Three dates sit close together and get written inconsistently across news reports, so here is what each one is. October 13 is the Cabinet decision, October 20 is promulgation, and April 21, 2027 is entry into force. From November 8, 2026, the date of writing, that leaves 164 days — a little over five months.

For a reader, who carries the obligation matters more than the dates. Paragraph 2 of the newly added Article 15-2 lists the obligated parties in separate subparagraphs, and heads of state agencies are not the only ones there. Entities carrying out projects on government contributions, subsidy recipients and indirect subsidy recipients, local subsidy recipients, and parties equivalent to these each fall in, within a scope set by Presidential Decree. That means a private company or a university lab can land inside the scope if the data was built with public funds. How far "the scope set by Presidential Decree" reaches has not been fixed yet; that waits on the amended enforcement decree.

The same amendment widened the door on the other side. While imposing a new duty on those putting data into the repository, it opened up the taking-out side by changing "the private sector" in the existing Article 15 to "anyone." State agencies, in other words, also become users who draw data from this repository. As the inflow grows and the set of recipients widens, the question coming up in the next section gains weight: who looks again at data once it has been uploaded, and when.

The duty comes with exemptions, and claiming one runs through a consultation procedure. What those exemptions are and why they are loose is examined in our earlier article on the exemption clause. This section looks only at scope, not exemptions.

A word on how far this article verified the statutory text. Article 13 of the enforcement decree was read directly on the National Law Information Center. By contrast, the subparagraph text of Article 15-2 in the amended act came by way of a law firm's commentary, because the promulgated text itself could not be opened. The article number also shows up in a different 2026 amendment, so anyone citing it is safer checking the promulgated text once more.

The window open right now, then, is the last voluntary one before the duty arrives. And that window did not close in August. Copied straight from the private-sector notice, the submission period reads as follows.

What is inside the parentheses changes how this article should be read today. The August notice is not old news but a desk still taking submissions. For anyone weighing whether to contribute data, this is usable information rather than a retrospective. The one caveat: as of November 8, 2026, no announcement of selection results is found in the notices.

Start with the statute. Saying there is no legal basis for quality would be wrong. Article 13, paragraph 1 of the AI Framework Act's enforcement decree lists six tasks for the integrated provision system, the fifth of which is "providing information related to the valuation and quality management of training data." Article 15-2 of the amended act adds measures for the stable quality management of data that has been connected and provided. The enabling clauses exist.

Three things the clauses do not say. How often the data gets looked at again (the interval), who stands behind the guarantee (the party), and how anyone who already took a copy is told (notification). Of the three, notification cuts deepest. AI Hub's data use policy spells out the user's side of the obligation.

The duty on whoever finds the problem to report it and delete the copy is there in so many words. The opposite direction — a clause saying re-verification results will be passed to people who already downloaded the data — is not found in the use policy. Yet AI Hub already has what it would take to do that. Getting data requires mobile identity verification, a per-dataset request, and approval. This is a platform that knows who took what.

What about the interval? It is absent from the Korean statutes, but other documents doing the same job put a number on it. The De-Identification Guidelines for Structured Data, issued by the Information and Privacy Commissioner of Ontario on October 15, 2025, is a 96-page revision that overhauls the 2016 edition; it says re-identification risk assessments should be repeated periodically because the risk environment shifts, and specifies a typical interval of two or three years. The same guidance cites a re-identification probability threshold of 5–9%, aligned with ISO/IEC 27559. Clinical trial data practice has its own convention of reassessment after 18 to 24 months. None of this argues the number must be two years. It argues that documents already exist that write the interval down as a number.

So how do national data platforms elsewhere call published data back in? Look into it and Korea is not the laggard. It is close to the only one in a position to do it at all.

Table 6. Whether a documented path exists for calling published data back in. The scope is what could be confirmed in public documents as of November 8, 2026.

The facts here are current as of November 8, 2026, and the personal-information re-verification contract wraps up in December. What shape those 150 results reports take then, and what state the four withdrawn datasets return in, are the next things to look at.

What this program is trying to buy for KRW 60 million per dataset is very close to what DataClinic does: diagnose quality, run de-identification, and write the deliverables — data specification sheets, metadata definition documents, labeling guides. So this article is not a spectator's look at someone else's procurement. It is a rare public record of the unit price and scope of work for the job we do. The four stages of the re-verification contract in particular — start from sample detection, fully inspect only the datasets that get flagged, apply additional de-identification by data type, then verify adequacy — read almost like a published reference architecture for a quality diagnosis pipeline. The requirement of a separate report for every single dataset even shows the shape the diagnostic output should take.

8.1Three lessons about the shelf life of a guarantee

First, de-identification is a judgment that expires. "De-identification complete" is a conclusion reached with the auxiliary information available at that moment and the detection technology of that moment, and both change with time. That is why the Ontario guidance writes a reassessment interval of two to three years.

Second, a guarantee attaches to a document, not to the data. Having a quality management report and the data being clean right now are two different propositions. And at this intake window, that report is a document collected "only where it can be submitted." So only the report remains, and no one can say whether what it recorded is still true.

Third, "synthetic" is a label, not a guarantee. The script may be invented and still be read aloud by a person, and the generative model itself may memorize and spit back the originals it trained on. All three turn up in every customer data review we run. They are not something that happens only on national platforms.

8.2Five questions to ask before a dataset leaves the building

Turn the shape of the failures seen here into questions and you get five. These are what a company should check before handing a dataset to an outside party, or before taking one in.

• When was this guarantee issued? Do the date of the ruling and the detection rules used at the time travel with the data?
• When is the next verification? Is there a set interval, or does the data only get opened after something goes wrong?
• How do you reach the copies already out there? Is there a list of who took one, and a path to reach them?
• Who stands behind the guarantee? When the builder, the verifier and the publisher are three different parties, who carries the liability?
• What substantiates the label "synthetic" or "pseudonymized"? Is there a record of how it was generated?

The third is the one most often empty. Without a list there is no way to undo a mistake once you find it, which makes this the most expensive gap in practice.

8.3Where Pebblous stands

Earlier articles on this blog mostly examined the moment of intake: is it a duplicate, how many datasets are there, may it be used, must it be handed over, does a standard exist. What this one looked at is what is already inside. The question was what becomes of data that reached the repository and went public once time passes, and who checks that state again.

Sources are grouped by kind. Items 1 through 8 are the notices, statutes and use policy opened directly for this article. Items 9 through 17 are the news coverage that supplied the dates and amounts. Items 18 through 28 are the academic literature and foreign norms on the shelf life of a guarantee. The last group is earlier Pebblous articles that looked at the same repository from other angles. Korean-language sources are cited by their English titles in translation; the quoted passages in the body are our translations of the Korean originals.

Notices, statutes and the use policy

• 1.AI Hub notice, "Guide to participating in the private-sector call for AI training data provision," August 3, 2026. aihub.or.kr (Korean). The submission-document wording in Section 3 and the submission period in Section 6 come from this notice.
• 2.Seoul National University College of Engineering notice, "Call for submissions on university-sector AI training data provision, registration and quality support," posted September 1, 2026. me.snu.ac.kr (Korean). The "on condition of release" wording in Section 2 and the submission method in Section 3 are here.
• 3.AI Hub, "Data Use Policy." aihub.or.kr (Korean). The two quoted sentences on user obligations in Section 7 come from this document. Checked November 8, 2026.
• 4.AI Hub, "AI Data Quality Management Guidelines," four volumes, published February 27, 2026, files reposted April 15, 2026. aihub.or.kr (Korean). The per-volume version numbers and the "three volumes in total" footnote in Section 3 come from here.
• 5.AI Hub dataset listing. aihub.or.kr (Korean). The denominator of 977 in Section 5 is as displayed on November 8, 2026.
• 6.Framework Act on the Development of Artificial Intelligence and Establishment of a Basis for Trust, Article 15, National Law Information Center. The enabling clause for the integrated provision system cited in Section 1.
• 7.Enforcement Decree of the same Act, Article 13 (Construction and management of the integrated provision system for training data). The quotation from paragraph 1, subparagraph 5 in Section 7 is here.
• 8.Amendment to the same Act, promulgated October 20, 2026, in force April 21, 2027, adding Article 15-2. ⚠️ The promulgated text could not be opened, so this came by way of the commentary at item 15. That limitation is stated in Section 6.

News coverage

• 9.Electronic Times (Etnews), Park Jong-jin, on the launch of the first full census of public-sector AI training data, April 10, 2026. etnews.com (Korean). The 100 datasets and KRW 6 billion in the Section 1 timeline come from here.
• 10.eNews Today, Baek Yeon-sik, on the AI Hub overhaul and the integrated provision system, February 10, 2026. The "100 registered and connected within the year" wording in the Section 1 timeline comes from here.
• 11.IT Daily, Yang Seung-gap, on the procurement notice for building the integrated provision system, July 2, 2026. The KRW 2.879 billion and 120 days in Section 1 are here.
• 12.Ministry of Science and ICT press release, on opening training data from the Sovereign AI Foundation Model project, August 27, 2026. korea.kr (Korean). The account of quality and personal-information verification in Section 4 follows the sentences in this release.
• 13.NewsPost, Kim Yun-jin, "[Exclusive] Citizens' names and phone numbers exposed in NC AI's Sovereign AI data … AI Hub release suspended," September 17, 2026. newspost.kr (Korean). The two sides' quoted explanations in Table 4 come from here.
• 14.ZDNet Korea, Lee Na-yeon, "Personal information in 150 datasets published on AI Hub to be examined again," September 29, 2026. zdnet.co.kr (Korean). The contract name, budget, four work stages and schedule in Section 5 come from here.
• 15.Law Times, Jung Ho-jung and Lee Gun-hee of Daeryook & Aju law firm, commentary on the amended AI Framework Act, November 5, 2026. lawtimes.co.kr (Korean). The categories of obligated parties in Section 6 came by way of this commentary.
• 16.Kyunghyang Shinmun and others, on the Board of Audit and Inspection's findings on public AI training data, August 12, 2026. khan.co.kr (Korean). The 908 datasets and KRW 1.6328 trillion in Sections 2 and 5 are that audit's tally.
• 17.Electronic Times (Etnews), on newly designated Data Safety Zones, May 8, 2025. etnews.com (Korean). Background for the Data Safety Zone passage in Section 2.

Academic literature and foreign norms

• 18.Gebru, T. et al., "Datasheets for Datasets," Communications of the ACM 64(12), 2021 (arXiv:1803.09010). The Maintenance questions in Table 3 come from this paper.
• 19.Subramani, N. et al., "Detecting Personal Information in Training Corpora: an Analysis." The post-release audits of C4 and The Pile in Section 5, and the 2.7% and 2.5% figures, are here. Both percentages take detected instances as the denominator.
• 20.Dodge, J. et al., "Documenting Large Webtext Corpora: A Case Study on the Colossal Clean Crawled Corpus," EMNLP 2021. aclanthology.org The C4 size figures in Section 5 follow this paper.
• 21."A Common Pool of Privacy Problems" (arXiv:2506.17185), 2025. The one-line summary at the end of Section 5 is this paper's conclusion.
• 22."When Privacy Isn't Synthetic: Hidden Data Leakage in Generative AI Models" (arXiv:2512.06062). The source for the structure-and-neighborhood route among the two synthetic-generator leakage paths in Section 4.
• 23."When Tables Leak: Attacking String Memorization in LLM-Based Tabular Data Generation," PoPETs 2026. petsymposium.org The verbatim-memorization attack in Section 4 is this paper.
• 24."SoK: Reconstruction Attacks on Synthetic Tabular Data" (arXiv:2606.08372). The TabDDPM 13.5pp and CTGAN 0.5pp figures in Section 4 come from this systematization paper.
• 25.Information and Privacy Commissioner of Ontario, De-Identification Guidelines for Structured Data: Updated and Expanded, October 15, 2025, 96 pages. The typical reassessment interval of two to three years and the ISO/IEC 27559-aligned threshold of 5–9% in Section 7 are here.
• 26."The risk of re-identification remains high even in country-scale location datasets," Patterns, 2021. Background for the claim in Section 7 that re-identification risk shifts with time and auxiliary information.
• 27.Congressional Research Service, Availability of Federal Data: Policy Considerations for Disclosure, Preservation, and Governance, R48889. The data.gov row in Table 6 reflects this report's observation.
• 28.Connecticut Open Data Retirement Guidelines and NYC Open Data Dataset Removals. The retirement-register examples in Table 6.

The same repository from other angles

• 29.The Government Built the Training Data. Then Other Agencies Built It Again. — on the Board of Audit findings that sit behind this overhaul.
• 30.Korea Opened Its Sovereign AI Training Data. Two Official Records Count It Differently. — the counting dispute deferred in Section 4 is handled here.
• 31.Government AI Data Must Go to AI Hub, Unless Its Maker Says No — on the exemption clause deferred in Section 6.
• 32.Korea's AI-ready data standards are scattered across catalogs and DataOps — the article this one parts ways with in Section 3. That piece drew the absence of standards; this one looks at the place where the documents exist but filing them is optional.
• 33.Korea Pulled 100 Public Datasets Forward a Year. The License Change Was a Different List. — the separate initiative covered by the fourth row of Table 2.