Constructing Language Judgments: Rubric Distillation and Evidence-Grounded Reasoning

Gnomon investigates the construction of language judgments through rubric-grounded reasoning, structured supervision and an evidence-verifying execution harness. Rather than asking a model to predict a score and justify it afterward, the architecture compiles declarative rubric specifications into a versioned item bank shared by evidence elicitation and scoring. Model-derived observations pass through exact-span verification, relation-level credit accounting and constrained criterion aggregation, making the transformations from source text to judgment inspectable. The learning infrastructure operationalizes rubric distillation through structured extraction targets, verifier-filtered attempt selection and candidate-bound critique-and-repair examples. Completion-focused parameter-efficient training preserves example weighting across microbatches; chunked vocabulary projection and supervised-token accounting support resource-aware execution. Together, these components connect the inference harness, judgment construction and training-data feedback loop. IELTS writing assessment is the implemented application, not a claim that the architecture has been validated across all language-judgment domains. The report documents system interfaces, five explanatory figures and an executed archived-input replay. Historical Task-2 records report criterion MAEs of 0.226–0.319 on 157 essays with source-specific calibration; these are contextual results, not a current-engine comparative benchmark. Reasoning refers to explicit evidence transformations, not access to hidden model cognition. Independent technical report, revision 14.