Thais Castello Branco argues that design, writing and other subjective tasks are difficult to train and evaluate because good results depend on audience, context and time. She distinguishes the parts of a task that can be checked against clear criteria from those that require human judgment, proposing that each part should use the method best suited to it.
Thais Castello Branco uses brand adherence as an example of making a fuzzy goal more verifiable. Decomposing a brand into colors, typography, spacing, motion and texture can give an evaluator concrete constraints. She argues that a new design should satisfy those constraints without having to copy the original, and cautions that model-based judges can introduce hallucinations or reward hacking.
Thais Castello Branco describes repetitive, generic output as a tendency toward an average that may be poorly suited to creative work. She proposes preserving information about whose preferences are being expressed and in what context, allowing distinct tastes to guide training rather than treating their disagreement as noise.
Thais Castello Branco emphasizes careful expert selection and specific feedback tied to the precise visual or code component being judged. In human quality assurance, she separates disagreements about measurable requirements from differences in style or aesthetics, arguing that the latter may represent useful preference data rather than a flaw.
Thais Castello Branco advocates quality over quantity when collecting data for subjective domains. Her framework prioritizes deliberate task design, contextual preference information and detailed expert reasoning, while acknowledging that such data takes effort and domain expertise to produce.
Watch on YouTube
