The load-bearing detail is the evidence Krafton keeps. Its tool captures source code, conversation logs, prompts, tool settings and traces of verification runs from a candidate's session [5], which is five classes of evidence where the conventional screen kept one, the submitted artifact [3][1]. Four stages get scored instead: problem definition and prioritization, the quality of AI collaboration and instructions, evidence-gathering and decision-making, and iterative improvement and error recovery [4], four scored process stages set against the single graded deliverable a conventional screen produced [4].
Krafton is candid about the weak joint. With an LLM, the same submission can return different results from one run to the next [7], so evaluation criteria and score conversion are controlled by code while the judgments that require reading context go to LLM agents [9], and a meta harness repeatedly tests and refines the grading agent's execution structure [8]. Fencing the scoring arithmetic in code is the right call, and it also means the score travels no further than the harness that produced it. The instrument itself was built using OpenAI's Codex [2], so the rig that measures how well you direct agents is itself a directed-agent product.
What Krafton is not doing with that engineering time is the part I would price. The AI Frontier Division built an internal hiring instrument roughly ten months after the October declaration that the company would become AI-first [10][2], and then took it outside the building, grading tasks set by CJ Olive Young at a hiring-linked hackathon at PUBG Seongsu in Seongdong [12], with plans to widen the tool's scope [15] and an ambition, in AI Frontier Division head Park Jae-min's words, to set a new standard for assessing AI-native talent [13]. The premise a Krafton official gives is that problem-solving ability does not show up in the output alone [14].
There are a few ways this plays differently. The process signal is real, and the hires ramp faster than portfolio-screened ones did. Or the four named categories become coachable, and candidates learn to narrate hypotheses at the grader rather than form them [4]. Or the rig is the deliverable [15][13], and hiring quality is a by-product for whoever eventually runs it on their own applicants. This is probably wrong, but I would weight the second, on the grounds that a published rubric decays into performance and this one's stages are already in print [4]. What would move me is dull evidence: retention or ramp numbers on Cofa-Probe hires against the previous cohort, or a variance figure showing two runs of one submission landing on the same score [7].
