Image capturing the concept of One Certificate, One Benchmark Line: Korean AI Ran Into the Limits of Static Proof Today An image capturing the core concept of this article.

Two pieces of paper wobbled on the same day

One is the direct-production certificate. Hold this piece of paper and you register as a domestic server maker on Korea’s public procurement platform, Nara Jangteo. According to DigitalDaily’s investigative series on server sovereignty, about 260 companies are registered this way. The other is a benchmark scorecard. According to an exclusive report from inews24, the Ministry of Science and ICT is reviewing how it scores the third round of its sovereign foundation model evaluation. In the second round, benchmarks made up 40 of 100 points, and suspicions have surfaced that participants were approached with offers to raise those scores.

The two cases sit in different industries under different ministries. But they fail the same way: proof issued at one point in time cannot vouch for what happens afterward. Read today’s news through that lens, and stories that looked scattered line up.

Paper doesn’t see what happens after assembly

The server side of this is especially blunt. Import a bare-bones server from overseas, plug in a CPU and memory domestically, and it counts as a domestic product. One company claims it designs its own motherboards, a capability it says only 20 companies worldwide have, but competitors counter that the core components are the same, so there is barely any technical difference. In fact, standard benchmarks like SPECjbb show almost no performance gap between domestic and foreign servers.

The harsher number comes next. Of the roughly 3 trillion won in annual public server budget, only 1 to 2 percent flows through procurement that favors domestic products; the remaining 98 percent defaults to foreign vendors bundled into systems-integration contracts. The system for issuing certificates is intricate, but most of the budget simply walks past it. This investigation shows, in numbers, what happens when sovereignty is defined by paperwork: paperwork is all you get more of.

To avoid a misread here: domestic assembly is not worthless. Having assembly lines and maintenance staff in-country is an asset in itself if a supply chain crisis hits. The problem is that this value and the value of design capability get stamped with the same single piece of paper. A label that doesn’t distinguish between the two gives no information, and a label that gives no information gets ignored by procurement officers. That 98 percent is the size of that dismissal.

A score is stamped once, but the model keeps changing

The benchmaxxing controversy is a different symptom of the same illness. Vice Minister Ryu Je-myung said no evidence of overfitting or memorization has been found. What matters is that the system is shaking even without confirmed wrongdoing. Participating companies are unanimously asking for the benchmark weight to be reduced, not because they distrust the score, but because they no longer believe the score represents what it was meant to measure.

Three companies made it to the third round of evaluation, LG AI Research, SK Telecom, and Upstage, and the ones that pass will each receive support worth up to about 1,000 NVIDIA B200 GPUs, roughly 120 billion won. Separately, the Everyone’s AI program allocates up to 512 B200 GPUs to selected operators. One line of scoring decides the allocation of over 100 billion won in resources. When a metric carries that much weight, it becomes a target for optimization. The old warning that a metric stops being a metric the moment it becomes a goal is replaying itself at national budget scale.

That said, jumping to “let’s drop the score” is the wrong conclusion. Without benchmarks, the review becomes a contest of persuasive presentation slides, and that is far easier to game. Even just among the three finalists, the strengths differ. Upstage leans on million-token long-context processing, SK Telecom on reasoning at the level of an International Mathematical Olympiad gold medal, LG AI Research on a 750-billion-parameter model. Comparing candidates on such different axes still requires a common scale. The real fix isn’t removing the score but making sure observation continues after the score is stamped. If what a selected team actually did with its GPUs, and how users actually used the result, keeps getting recorded, the weight carried by a single scoring event gets that much lighter.

The drug ministry raised this issue first

Interestingly, today’s most forward-looking answer came not from the AI industry but from a regulator. Son Mi-jung, a division head at the Ministry of Food and Drug Safety, diagnosed that current regulation, built around fixed machine-learning models, cannot handle generative, continuously learning, and agentic AI medical devices. She proposed three redesign axes: data utilization, handling continuous learning, and securing reliability and accountability.

The core of this diagnosis is that the moment of approval and the moment of real-world use pull apart. A system that keeps learning does not stay in the model it was approved on. Approvals for digital medical devices carrying AI software are projected to rise from about 500 last year to 700 to 800 by the end of this year, and simply reviewing more cases won’t keep pace. This leads to a conclusion: shift from one-time certification to lifecycle-long observation. Server procurement and model evaluation are stuck at exactly the same point.

The Everyone’s AI program structure, also revealed today, reads the same way in this context. Six operators and consortiums, including SKT, Kakao, and KT, entered the bid, splitting along different axes: SKT full-stack, Kakao consumer touchpoints, KT multi-LLM. One notable requirement is that at least 50 percent must be domestic models meeting sovereign foundation model criteria, with at least 30 percent from other companies’ domestic models. With beta running through late September and full service planned for December, this program will be decided in real-world use, not in the review room. Domestic AI procurement is already shifting its weight from document review to operational track record.

Traces of proof descending into execution records

So what will replace the certificate? Today’s remaining stories drop hints. The common thread is that the grounds for proof are becoming execution logs, not documents.

At SK Ecoplant, a field engineer used vibe coding to build an agent that auto-summarizes a 1,600-page geotechnical survey report and visualizes it in 3D. The company runs a three-stage program spanning AI adoption, capability certification, and agent development and productization, and about 200 employees have completed the certification track. Here, what proves the tool’s value isn’t an adoption report but the pile of reports the agent has actually processed.

The security side is even more blunt. Swimlane launched a feature that automatically classifies alerts into three paths: deterministic automation, AI-assisted investigation, and fully agentic investigation. One healthcare customer routed only the top 10 percent of over 180 daily threats into agentic investigation and cut costs by 90 percent. CEO Cody Cornell was direct that at hundreds of thousands of alerts, five to ten dollars in token cost per alert is unaffordable. Because a record persists of which alert cost how much, the savings claim becomes a fact rather than an assertion.

The KHF 2026 exhibition floor pointed the same direction. A reading-support service that auto-generates preliminary findings for chest X-rays appeared, a digital pathology viewer that’s already deployed at Seoul St. Mary’s Hospital is set to expand to seven hospitals, and autonomous robots that carry medications and specimens were on display too. The moment software crosses into the physical world, the accountability problem the drug ministry flagged can only be resolved by a record of what that robot did that day, not a copy of a certificate. If there was a night when a fall-detection camera didn’t sound an alarm, the answer lives in the decision log from that moment, not in a certificate.

Cost, meanwhile, doesn’t wait for the debate to settle. According to a report citing Reuters, Samsung Electronics raised prices for new 4nm and 5nm orders by 10 to 15 percent, and forecasts point to older DRAM prices jumping 50 percent as demand piles into HBM. The infrastructure bill keeps arriving monthly even while evaluation criteria are being fine-tuned. This might look unrelated to the discussion above, but it isn’t. As costs rise, organizations need grounds to explain where and how much was spent, and those grounds ultimately come from records kept at the level of individual executions. The more expensive things get, the more auditability shifts from a virtue to a requirement.

What to leave behind instead of a certificate

To sum up, today’s question is this: how do you turn a certificate that’s issued once and done into evidence that keeps accumulating.

This is the premise ThakiCloud held onto when building Paxis. Paxis is an Agent Native Cloud, currently at v1.1 GA as a shipping product. In this product, Skills, Tools, Policies, and Audit Logs are not add-ons, they are first-class resources. Every time an agent runs, a record is kept of which skill it used, which tool it called, and which policy gates it passed. Autonomy is managed in stages from L0 to L3, so riskier work can be routed through human approval. Execution happens in an isolated sandbox, and where sovereignty requirements are strict, it can run directly on an on-premises Kubernetes cluster. There’s also a path for picking the model per task, so the pattern Swimlane demonstrated, spending expensively only on the top 10 percent, can be applied across an organization’s full range of work. Through MCP connectors and a skills marketplace, business teams can assemble their own skills for their own work, and as with the SK Ecoplant case, the more the people building tools shift from IT departments to the field, the more control and recordkeeping matter. Anyone can build, but what they did has to be kept on record.

Cloudera’s Anywhere Cloud, launched yesterday, promises to unify public, on-premises, sovereign infrastructure, and edge under a single console, and it’s aiming at the same demand. The approach is to run workloads where the data already sits, without moving it. This means more competition, but it also means the market is converging on what enterprises actually want.

Certificates will still be needed going forward. But the fact that they’re not enough on their own showed up in two pieces of paper on the same day. Just as a direct-production certificate can’t fully explain the assembly process, a benchmark score can’t fully explain how a model is actually used. What remains is the execution record. If a record survives of who did what with which authority, it can be audited later and used as the basis for the next evaluation. If the sovereign foundation model evaluation, which will pick the final two teams early next year, shifts weight toward user evaluation, that direction will end up pointing at the same place.

Closing out today’s news, one sentence remains: proof is not issued, it is accumulated. Whether it’s servers, models, or agents, the organizations that earn trust next quarter will more likely be the ones that can show what they’ve actually done, not the ones with better paperwork. That preparation doesn’t need to wait for the system to change. It can start today.

References

This article was compiled from the following news sources.

Tags: agentops, enterprise-ai, paxis, thakicloud

Categories:

Updated:

You may also enjoy