Large-scale annotated datasets underpin progress in natural language understanding, yet human annotation remains a persistent bottleneck. Large Language Models promise to bridge this gap, but their reliability as annotators is poorly understood. We evaluate 14 open-weight instruction-tuned models and 5 matched base–instruct pairs across 51,189 instances spanning 10 heterogeneous datasets, revealing that annotation failures are structurally predictable rather than random. We identify three recurring failure modes: Safety Hallucination, Intensity Collapse, and Entailment Bias. Safety hallucination can precede alignment rather than being introduced by it: some base models flag up to 100% of benign text as hate speech. Alignment benefits scale inversely with base model capability, yielding a +21.6pp gain for the weakest base model yet a negligible +0.5pp shift for the strongest. KL divergence between base and instruct distributions is 2.0x higher on incorrect predictions, providing a lightweight instance-level signal for alignment-associated errors. Inter-prompt unanimity achieves 72–94% accuracy at 71–90% coverage on clean tasks, yet can actively degrade performance on tasks dominated by shared systematic bias. We translate these findings into a practical deployment framework with pre-screening diagnostics and tiered quality-routing policies, establishing principled boundaries for when LLM annotation can be trusted and when human intervention remains necessary.

Pretrained language models have transformed text classification, yet their computational demands often render them impractical for resource-constrained settings. We propose a linguistically-grounded framework for context minimization that leverages theme-rheme structure to preserve critical classification signals while reducing input complexity. Our approach integrates positional, syntactic, semantic, and statistical features, guided by functional linguistics, to identify optimal low-context configurations. We present a methodical iterative feature exploration protocol across 6 benchmarks, including our novel CMLA11 dataset. Results demonstrate substantial efficiency gains: 69-75% reduction in GPU memory, 81-87% decrease in training time, and 82-88% faster inference. Despite these resource savings, our configurations maintain near-parity with full-length inputs, with F1 (macro) reductions averaging just 1.39-3.10%. Statistical significance testing confirms minimal practical impact, with some configurations outperforming the baseline. SHAP analysis reveals specific feature subsets contribute most significantly across datasets, and these recurring configurations offer transferable insights, reducing the need for exhaustive feature exploration. Our method also yields remarkable data compression (72.57% average reduction, reaching 92.63% for longer documents). Ablation studies confirm synergistic feature contributions, establishing our context minimization as an effective solution for resource-efficient text classification with minimal performance trade-offs.