GraphProof-QA Direct Adapter (System A)

LoRA adapter (r=16, alpha=32, 3 epochs) teaching Qwen2.5-1.5B-Instruct to answer MetaQA movie questions directly. Baseline for the compile-then-execute comparison — same base, same data budget as the DSL adapter.

Training: 30,000 question/answer pairs (10k per hop). See graphproof-qa.

Results (2,000 test questions/hop): hop1 14.2%, hop2 22.2%, hop3 66.3% — 34.2% overall. Memorizes; hallucinates plausible entities on unseen questions (5.8% on renamed entities vs 81.4% for the DSL system).