NCAA men's lacrosse rulings depend on exact timing, location, exceptions, and version-specific language, which makes the task difficult for a general-purpose language model that may rely on incomplete parametric knowledge or confidently collapse distinct conditions. I built a lightweight retrieval-augmented generation system that retrieves structured records derived from the official NCAA men's lacrosse playing-rules resources, expands the governing rule family and linked companion provisions, and sends that evidence to Qwen2.5-1.5B-Instruct. RAG improved LegalBench character Recall@5 from 0.000 to 0.464, TechQA token F1 from 0.265 to 0.335, and RAGTruth hallucination F1 from 0.104 to 0.216. The held-out NCAA evaluation also exposed the central limitation: retrieving the decisive record did not guarantee that the small reader would apply every condition correctly.

The evaluated NCAA corpus contains 42 independently retrievable records organized into 16 rule families. Each record stores a rule or approved-ruling identifier, citation, title, record type, source, printed page, source text, search concepts, authority rank, and explicit companion-record links. The source materials were the official 2025–26 NCAA men's lacrosse rules book and the 2026 video-review parameters; the original PDFs and exact-text corpus are not redistributed in this public repository. Twenty scenario-based NCAA examples were constructed for pipeline evaluation, with 18 used for retriever development and ex_001 and ex_010 held out until final evaluation. The external suite used all 776 queries in LegalBench-RAG-mini, all 314 TechQA test examples from RAGBench, and 900 question-answer examples derived from RAGTruth.

The selected pipeline uses intfloat/e5-small-v2 with its recommended query: and passage: prefixes, normalized 384-dimensional embeddings, and cosine retrieval implemented as inner-product search in FAISS. It retrieves the top five records, fully expands the strongest retrieved family, adds linked companion records, and then adds a second retrieved family if space remains, with a maximum of 14 context records. The grounded prompt instructs the reader to preserve timing, location, and negation; cite only supplied records; and return a ruling, explanation, citation, and uncertainty statement. Generation uses do_sample=False, repetition penalty 1.05, and a normal NCAA limit of 288 new tokens. E5 was selected over normalized BGE cosine and unnormalized BGE Euclidean retrieval because all three achieved perfect family and decisive-record Hit@5 on the 18 development questions, while E5 produced the strongest ranking metrics.

LegalBench-RAG-mini measures precise legal retrieval, RAGBench-TechQA measures retrieval plus technical question answering, and RAGTruth-QA measures whether a reader can identify unsupported or contradictory claims in retrieved context; the two held-out NCAA scenarios assess grounded application of the project corpus. The selected system is compared with the same Qwen2.5-1.5B-Instruct reader without retrieval, Qwen2.5-0.5B-Instruct with identical frozen evidence to test reduced capacity in the same model family, and Llama-3.2-1B-Instruct with identical frozen evidence to provide a similarly sized cross-family comparison. Qwen2.5-1.5B with RAG was strongest on TechQA and produced the most balanced overall result. Qwen2.5-0.5B had the highest RAGTruth F1 but achieved it with very high recall and poor precision and accuracy. Llama achieved the highest automatic NCAA ruling overlap, yet manual review found no fully correct NCAA response from any system.

LegalBench is retrieval-only, so the three RAG readers share the same selected E5 score. Eight Qwen2.5-0.5B TechQA answers reached the 768-token cap but remained scorable. Four of Llama's 900 RAGTruth classifications required a documented deterministic allowed-label log-likelihood fallback.

The system is intended for coaches, players, officials in training, analysts, and educators who want an evidence-grounded starting point for discussing NCAA men's lacrosse scenarios. It is best suited to learning, post-game review, and rule-reference workflows where a human can inspect retrieved evidence. It is not an authoritative officiating tool and should not determine a live-game or disciplinary outcome without review by a qualified official and the current official publications. The public repository includes the inference code but does not redistribute the third-party exact-text corpus; a user must supply an approved local artifact directory containing the three files documented in artifacts/README.md.

The prompt contains a fixed system instruction, complete structured evidence records selected by retrieval and rule-family expansion, and the scenario to be decided.

The model should return four labeled sections in plain text. The uncertainty section must identify missing material facts rather than silently assuming them.

The NCAA held-out evaluation contains only two scenarios and cannot establish broad officiating reliability. More importantly, all three RAG readers received the same retrieved evidence and still failed substantive review on both NCAA cases: models reversed explicit conditions, omitted a branch of a two-part question, or failed to acknowledge a missing material fact. Automatic token overlap therefore overstated rule correctness. The corpus covers a deliberately structured subset of the current rules rather than every possible NCAA provision, and it must be rebuilt or reverified when the rulebook changes. The evaluation used an RTX A6000 rather than a client device, so the project demonstrates a lightweight reader architecture rather than completed edge-device benchmarking. The public repository does not include the third-party exact-text corpus or its derived runtime index, and the system is not affiliated with the NCAA or a replacement for the current official rules or human judgment.

The selected Check-In 4 experiment ID is e07ab55131b7297d; the final comparison experiment ID is c777152194b9eba5. Full counts, source hashes, and comparison disclosures are stored in results/reproducibility_manifest.json.