The Claim as Circulated

The assertion describes a reverse-built evaluation. Instead of testing a model on questions written independently of its outputs, the test set was assembled backward: the desired answer was fixed in advance, and the question was written around it. A model graded against such a set can record a 99.8% success rate without ever demonstrating the ability to solve genuinely unseen problems — the score measures recall of a predetermined output, not reasoning.

What the Review Established

Reporters reviewed five primary sources for the claim. None was paywalled, geo-blocked or otherwise restricted; the "blocked — direct capture needed" flag was marked not applicable because every source opened normally. The core elements — the answer-first, question-later construction and the 99.8% figure — were corroborated across the reviewed material. The verdict falls short of full confirmation, however, signaling that elements surrounding the central claim were not pinned down to the same standard as the mechanism and the number itself.

Why It Matters for Korea's AI Sector

Benchmark inflation is a recurring hazard as Korean developers of large language models compete for visibility against global rivals. An evaluation written in reverse can be presented to investors, regulators and the press as a performance result, and near-perfect scores of this kind invite exactly the skepticism this review applied. The distinction between a model that solves problems and a test assembled to match a model's outputs is the line between engineering and marketing — and the sources indicate that, in this case, the test came second.

On the strength of five freely accessible primary sources, the claim that a 99.8% success rate was achieved by writing the answers first and attaching the questions afterward is Mostly True.