How does automating junior work destroy verification oversight?

The paper names two of "four structural failures that systematically erode the conditions for safe verification" under a "human-in-the-loop equilibrium" it calls inherently fragile, because "agentic labor scales elastically with compute and data" while "human verification remains inelastic—bound by the slow, biological accumulation of embodied experience." The first, "the Missing Junior Loop," holds that "by automating entry-level measurable work (Tm), we destroy the training ground required to build future experts (Snm)," so that "without involvement in execution, the pipeline of senior verifiers with the intuition to properly check the agents dries up." The second, "the Codifier's Curse," holds that "as experts apply their [non-measurable capacity] to verify tasks, they inevitably generate labels that digitize their tacit intuition into new" proprietary knowledge, so that "in doing so, they fuel the very automation that displaces them." The excerpt breaks off before naming the other two failures.

Both follow from the paper's premise that non-measurable expertise (Snm), the capacity needed to safely oversee AI, is "built through the friction of routine execution" rather than taught in the abstract. Automating that routine execution does not just remove jobs; it removes the mechanism by which the next generation of overseers would have learned to oversee. The Codifier's Curse compounds this at the senior end: the act of making an expert's judgment explicit enough to check an agent's output converts tacit skill into a reusable label, which becomes training data or proprietary knowledge that narrows the pool of tasks future juniors could have learned on. Read together, the two describe a verification pipeline that consumes its own inputs.

This gives a concrete mechanism to the institutional claim in What makes accountable judgment scarce when AI cognition is cheap?, which argues institutions "must preserve the learning, question-selection, authority, and bargaining conditions" for judgment to stay valuable — the Missing Junior Loop names exactly the apprenticeship channel that erodes if they do not. It is also a labor-level instance of the claim in Does incremental AI replacement erode human influence over society? that less human labor means less implicit alignment: here the labor removed is specifically the entry-level work that trained future verifiers, so the loss compounds rather than simply subtracts. It sharpens Does automation raise or lower the skills that remaining work demands? by giving a reason the remaining tasks' expertise requirement can rise even as the supply of people able to meet it falls: the training tasks are the ones being removed. The companion note, What becomes scarce when AI execution grows cheap?, states the macro argument these two failures feed.

The excerpt presents these as two items in an enumerated set the paper states has four members, so the full taxonomy of failure modes is not available here, and neither failure is backed by data on any actual workforce, firm, or occupation — both are derived from the paper's formal model of time allocation and human capital formation. What follows at that strength is a hypothesis worth checking empirically, that entry-level automation measurably shrinks the pool of qualified verifiers over time, not a demonstrated trend.