Imagine you're a restaurant health inspector. You have a checklist: food temperatures, handwashing stations, expiration dates. Now imagine someone builds a robot that checks every item on that list faster and more consistently than you can. Does that mean the inspector is redundant? No — because the inspector also notices that the kitchen staff seem rushed and resentful, that the walk-in cooler door doesn't seal properly even though the thermometer reads fine, that the new menu items don't match the equipment on hand. The checklist captures the measurable outputs. The inspector's real value is the integration across those outputs plus everything that never made it onto the checklist. That's the structural analogy at the heart of this critique. The target paper, "The End of Code Review," makes a clean four-function decomposition of peer code review: defect detection, style enforcement, knowledge transfer, and awareness. It then argues that LLM-based coding agents can execute each function at lower cost and higher throughput, therefore human review is no longer necessary. The logic is tidy. It is also, this rebuttal argues, fundamentally incomplete — because the decomposition itself is the error. The rebuttal identifies seven categories of reviewer contribution that survive outside the four-function frame. A reviewer's confusion is itself a finding — it signals that the code is too complex or the abstraction is wrong, and an LLM that always "understands" the input cannot produce that signal. Reviewers challenge whether a change should exist at all ("This solves the symptom, not the problem"), notice what is absent rather than what is present (absence blindness being a known LLM weakness), calibrate scrutiny based on who wrote the code and what context surrounds it, and engage in bidirectional sensemaking that changes both participants' mental models. None of these map to detection. The operational-context argument is particularly sharp. "We just had an incident in this service last Tuesday." "Legal told us not to log this field anymore." Human reviewers carry organizational state that lives outside repos, tests, and documentation. The target paper assumes the codebase is the complete context. It never is. This is not a gap that better retrieval-augmented generation will close — it is tacit knowledge that surfaces only when triggered by a specific diff in a specific moment. The accountability point cuts deeper than the target paper acknowledges. Being personally responsible for an approval shapes the quality of the review. An agent that signs off bears no consequences and has no incentive structure that fuels earnest evaluation. The target paper redirects this concern to "requirements engineering and post-deployment monitoring," which the rebuttal correctly identifies as hand-waving — pushing governance downstream without explaining who owns it there. The meta-argument is the strongest move. The rebuttal names the pattern explicitly: decompose human work into measurable functions, show the machine replicates each function, declare the human redundant. This is the substitution myth, and it fails at the same point every time — the human contribution that mattered most was the integration across functions and the adaptive response to unplanned circumstances. The original decomposition was never complete; it was merely the portion that was measurable. This is not an anti-AI argument. It is a precision argument about what code review actually is. If you define it as detection, agents win. If you define it as coordination, sensemaking, and governance with detection as a component, the picture changes entirely. The target paper chose its definition to fit its conclusion. This rebuttal refuses the framing.
