We introduce a corpus for Systemic Functional Linguistics (SFL) Appraisal analysis in Chilean Spanish, together with TULIO, a BERT model adapted to Chilean Spanish through continual pre-training. The corpus contains 2,546 texts annotated for the Attitude subsystem, with a focus on negative Judgment, and supports two span-level sequence labeling tasks: Attitude Types, which distinguishes Affect, Judgment, and Appreciation, and Judgment Subtypes, which identifies five finer-grained forms of social evaluation. To reduce the boundary ambiguity reported in prior Appraisal annotation work, we use a two-stage protocol that first marks structural spans and then assigns Attitude labels. Three expert annotators reach almost perfect agreement at the general level (Krippendorff’s α = 0.834) and substantial agreement for Judgment Subtypes (α = 0.647). We benchmark supervised sequence labelers and structured LLM agents. The highest observed F1 scores are 0.554 on Attitude Types from a Flair BiLSTM-CRF with TULIO embeddings and 0.410 on Judgment Subtypes from Qwen 3.6 Plus. Under Twitter-only training, supervised performance is genre-sensitive, with higher F1 on Twitter than on other genres. All model scores remain below expert agreement.