Korea picked three sovereign model teams. Publish the scorecard, not just the winners

THE SIGNAL IN ONE SENTENCE

South Korea ran a government-backed competition among four foundation-model teams. Three advanced. One did not. That sounds like a leaderboard story until you look at the team that missed the cut. Motif Technologies built Motif 3, a 314-billion-parameter mixture-of-experts model with 13 billion active parameters. South Korea's Ministry of Science and ICT says it recorded the strongest Artificial Analysis Intelligence Index performance among models from countries outside the United States and China. The ministry also praised Motif for building its architecture, tokenizer, optimizer and kernel in house. Motif still did not advance. Upstage, SK Telecom and LG AI Research did. The ministry says the final margin was narrow and the decision came from a 100-point framework: 40 points for benchmarks, 35 for expert review and 25 for user evaluation. The plain signal is this: when public money turns model development into a national competition, publish the complete scorecard, not just the names moving forward. The result is not necessarily contradictory. A team can lead one benchmark and trail on usability, ecosystem plans, deployment evidence or another part of the evaluation. That is one reason a multi-dimensional assessment can be more useful than a single ranking. But the ministry's public release gives averages, gaps and qualitative praise instead of the team-by-team component scores needed to reconstruct the decision. That leaves the most interesting public question unanswered: what evidence outweighed Motif's technical result? The four Phase 2 models were Motif 3 from Motif Technologies, Solar Open2 from Upstage, A.X K2 from SK Telecom and K-EXAONE 2.0 from LG AI Research. The ministry describes all four as openly released and lists different technical ambitions for each. Motif 3 emphasizes agents and coding. Solar Open2 is a 250-billion-parameter mixture-of-experts model with 15 billion active parameters, a stated context window of up to one million tokens and a hardware target of two NVIDIA H200 chips. A.X K2 is listed at 688 billion parameters with 33 billion active and is positioned around Korean-language, mathematics and industry use. K-EXAONE 2.0 is listed at 750 billion parameters with 37 billion active, support for ten languages and an Apache 2.0 license. These are ministry descriptions, not independent findings by The Plain Signal. A parameter count is not a usefulness score. A benchmark rank is not a deployment certificate. An open release also needs a license, weights, code, documentation and reproducible evaluation before outsiders can tell what is actually reusable. The evaluation design is more interesting than the model marketing. The benchmark portion carried 40 points. Twenty-five came from the Artificial Analysis Intelligence Index, or AAII, using nine tests across agents, coding, general capability and scientific reasoning. The ministry says Artificial Analysis ran that portion directly. The other 15 benchmark points came from a National Information Society Agency evaluation covering mathematics, knowledge, long-text comprehension, instruction following, Korean language, safety and reliability. The ministry says some questions and answers remain confidential to reduce contamination. Confidential tests can be valuable. If every answer is public, teams can optimize for the test instead of the capability. The tradeoff is reproducibility. Outsiders cannot fully audit a secret benchmark, so the public record should include its construction process, sampling rules, scoring method, security controls and independent validation without exposing the protected items. Across the four teams, the average benchmark score was 22.5 out of 40. The gap from first to fourth was four points. The expert review carried 35 points. Ten external specialists in algorithms, services and data evaluated team submissions over roughly one week, with question-and-answer sessions alongside the review. The ministry says the reviewers had no conflicts of interest. Ten points covered development strategy and technology. Ten covered development outcomes and future plans. Fifteen covered ripple effects and contribution plans. This is where the competition stops being only a test of model capability. A quarter of the full score effectively depended on plans for adoption, public contribution and ecosystem effects. That may be a sensible use of public support. It also makes evaluator instructions, conflict checks and scoring anchors essential public records. The expert-review average was 28.8 out of 35, with a 2.4-point gap between first and fourth. The last 25 points came from users. Forty-nine professional AI users, including startup chief executives, participated in a 15-point professional evaluation. The ministry selected 200 citizens for a 10-point general evaluation, and 185 ultimately participated. Each group received prompt guidelines and used absolute evaluation, meaning a model was rated against defined standards instead of being directly ranked against the others. The average user score was 17.6 out of 25. The gap between first and fourth was five points, the largest spread reported among the three major components. That could have decided the round. We cannot tell from the release because it does not publish each team's scores. This is why the scorecard matters. Start with one row per team and one column for every scored component. Show raw and weighted benchmark scores. Show expert scores for strategy, outcomes and ecosystem contribution. Separate professional users from general citizens. Add uncertainty intervals where repeated runs or sampled raters produce variation. Then publish the instructions. What tasks did professional users receive? Which services were citizens asked to try? Were interfaces equally polished? Did participants know the model names? Did every model have the same latency, tool access and safety settings? How were failed requests counted? Were accessibility needs included? Was Korean-language quality tested across formality, dialect and specialist domains? A usability score can measure the model, the product wrapper or both. The public should know which. The same rule applies to expert review. Publish the scoring rubric, reviewer expertise, recusals and conflict-screening process. An assertion that ten reviewers had no conflicts is useful. A documented procedure is stronger. The goal is not to expose private deliberation or invite harassment. It is to let the public see how judgment was disciplined. Next, connect selection to support. The ministry says GPU support for advancing teams will grow from roughly 768 NVIDIA B200 units in the first half of 2026 to about 1,000 in the second half. That is a substantial public resource. A scorecard should therefore include the compute already received, the compute promised next, milestones, energy reporting, release obligations and conditions for continued support. It should also explain what happens to the team that exits. Motif's technical work does not become worthless because it missed a narrow composite cut. The ministry explicitly calls the team's trial-and-error data and sovereign technology stack valuable national assets. If open ecosystems are the goal, the government should say which artifacts remain public, how outside researchers can use them and whether a strong non-advancing team retains access to shared infrastructure. The practical value goes beyond this competition. Public agencies around the world are funding domestic models, compute clusters and national AI champions. Many will face the same problem: one team leads a benchmark, another has the best deployment, a third offers stronger safety evidence and a fourth builds the most reusable public infrastructure. There is no universal formula for choosing among them. There can be a universal expectation that the formula and results are visible. For procurement teams, the scorecard can separate raw model capability from product usability and deployment readiness. For universities, it can show which weights, data notes and evaluation harnesses are actually available. For startups, it can reveal where public compute is concentrated and what contribution obligations come with it. For citizens, it can show why one team receives continued support while another does not. The cautions need to stay close. The ministry's release is both the official record and an argument for the program. Its global rankings depend partly on AAII, whose benchmarks change over time. The ministry itself warns that evaluation criteria changed between January and August, so a smaller score gap does not by itself prove a clean like-for-like improvement. The program's definition of sovereign AI also deserves inspection. Domestic architecture, data control, chips, licenses, cloud infrastructure and operational independence are different forms of sovereignty. A large Korean model running on imported accelerators can still create local capability, but the exact dependencies should be named instead of folded into one patriotic label. Selection is not proof of safety, social benefit or commercial success. User evaluation is not a substitute for red-teaming. Expert praise is not a production incident log. An open model is not automatically affordable to run. The ministry has built a genuinely richer evaluation than a one-number race. That deserves credit. It also creates a higher disclosure obligation. Publish the team-level points. Publish the prompts and scoring anchors. Publish the conflict controls. Publish the deployment evidence. Publish the compute bargain. The public paid for more than three winners. It paid for the lesson hidden in all four rows.