Much of ancient Greek literature has survived in a fragmented and often damaged state, whether due to centuries of manuscript-copying errors or the physical deterioration of texts originally written on papyri, stone tablets, or shards of pottery. Traditionally, restoring these lost or corrupted segments has relied on the specialized knowledge of Classical philologists. Recent advances in machine learning, however, make it possible to train language models on the corpus of extant Greek texts to propose plausible restorations in uncertain situations, thereby significantly aiding philological and textual scholarship (Assael et al., Nature 2022; Cullhed, Digital Scholarship in the Humanities 2025). In two preceding NAISS projects, I instruction-tuned pretrained causal language models to restore missing text in ancient Greek inscriptions and documentary papyri, establishing new state-of-the-art results, and conducted pilot experiments extending this approach to literary texts. These pilots confirmed that a compound approach—combining fine-tuned language models with metrical and stylistic filtering—is feasible, but also identified key limitations: single-pass token prediction cannot satisfy the multiple simultaneous constraints that literary restorations demand, and models need access to philologically curated training data that includes text-critical apparatus. This continuation project addresses these limitations by developing a compound AI system for restoring fragments of Athenian tragic dialogue in iambic trimeter. The system combines four components: (1) retrieval-augmented generation using a ColBERT index over the surviving corpus of Greek tragedy to supply parallel passages as context; (2) chain-of-thought instruction tuning, training the model to reason explicitly through metrical, dialectal, and stylistic constraints before proposing a restoration; (3) metrical filtering using a macronizer and automatic scansion tool for iambic trimeter; and (4) stylistic reranking using a classifier trained to distinguish among the three Attic tragedians. The project will also develop a comprehensive benchmark for evaluating language models on core philological tasks in Ancient Greek and curate semi-synthetic training data by augmenting base texts with variant readings and conjectures from the text-critical tradition. All models, datasets, benchmarks, and code will be released openly.