
Bare tokens vs real code: an 11x recall jump in LLM memory verification (3400 calls, 14 models)
14 models, 4 arms, ~3400 API calls. No model balanced false-accept and recall with pattern strings. With real file snippets: recall 0.08→0.88 at FA=0.02. The problem was the evidence, not the model.
Aug 14, 202610 min read12 reactions4 comments
