Large language models are increasingly used in evaluative tasks where they judge actions,attribute responsibility, and characterize people in morally charged situations. This thesisexamines how such systems exhibit gendered patterns in moral judgment when identicalinterpersonal scenarios are presented with different gender identity descriptions. It de-velops WAW-Bench, a controlled diagnostic benchmark of 1,200 scenarios spanning foursocial context domains (family, friendship, romantic relationship, workplace) and threemoral valence subsets (clearly moral, clearly immoral, ambiguous), in which the focalaction is held constant while the gender identity description of the focal character is var-ied across seven versions: Neutral, Woman, Man, Cisgender Woman, Cisgender Man,Transgender Woman, and Transgender Man. Model judgment is measured along threestructured dimensions — numerical Rating score, numerical Responsibility score, andopen-vocabulary Trait word projected into the warmth–competence space of the Stereo-type Content Model — and analyzed through a three-family comparison framework thatdecomposes identity-based variation along a gender axis (woman vs man), a markednessaxis (marked vs unmarked, within gender), and a Neutral baseline. The analysis iden-tifies three main findings. First, the Women-Are-Wonderful effect is replicated acrossall three measurement dimensions, with woman-labelled subjects receiving substantiallyhigher Rating scores, lower Responsibility scores, and warmer Trait-word coordinatesthan man-labelled subjects, particularly in ambiguous scenarios, and with the gap drivenby elevated evaluation of woman-labelled subjects rather than depressed evaluation ofman-labelled subjects. Second, the markedness axis operates as a distinct source of varia-tion: transgender-identity markers produce substantial leniency shifts that are symmetricacross genders and invisible to woman–man comparisons by construction; cisgender mark-ers do not. Third, the same identity descriptions produce coordinated shifts across allthree measurement dimensions, indicating that the bias operates as a structured pat-tern of differential moral interpretation rather than as isolated output variations. Thethesis contributes both a methodological framework for diagnosing identity-based vari-ation in LLM moral judgment and empirical evidence that this variation is structured,multi-dimensional, and not adequately captured by single-axis fairness measures.