LLM-based evaluators of natural language generation (NLG) quality are widely deployed as scoring tools and as automated training signals, yet the internal procedure by which they assign a rating remains poorly understood. We investigate this procedure mechanistically through an eight-attack perturbation taxonomy across the Readability and Adequacy dimensions of NLG quality, a generation pipeline that produces paired clean and corrupt summaries with controlled error intensity and explicit token-level modification maps, and a four-experiment battery of causal tracing, logit-lens vocabulary projection, and attention-head knockout applied to Themis (Llama-3-8B) and Prometheus (Mistral-7B). Both evaluators implement a structured, coherent evaluation pipeline operating in two stages: below layer 15, attention performs local error comparison and routes the result to the final input position; above it, the MLP cascade integrates the signal and writes the rating, with the decision crystallizing in the residual stream at a sharp late layer (L = 26 on Themis, L = 25 on Prometheus). Furthermore, a base-model control at the same scale (Llama-3-8B) reproduces the routing architecture and crystallization but not the stage separation, isolating the two mechanisms that fine-tuning specifically installs, suppression of below-L15 MLP contribution at the last position and a two-layer advance of the crystallization depth, indicating that fine-tuning sculpts an existing substrate rather than building the pipeline from scratch. We release the source code and data at https://github.com/himil-v/judge-mech.

This paper describes the contribution of team PHG to the WASSA 2022 shared task on Empathy Prediction and Emotion Classification. The broad goal of this task was to model an empathy score, a distress score and the type of emotion associated with the person who had reacted to the essay written in response to a newspaper article. We have used the RoBERTa model for training and top of which few layers are added to finetune the transformer. We also use few machine learning techniques to augment as well as upsample the data. Our system achieves a Pearson Correlation Coefficient of 0.488 on Task 1 (Empathy - 0.470 and Distress - 0.506) and Macro F1-score of 0.531 on Task 2.