CoreLine synthetic document extraction exercise All documents are authored synthetic examples, not client records. This is not a production readiness assessment or a model recommendation. Requirements: Python 3 and an existing local Ollama installation with qwen2:0.5b and llama3:latest already installed. The runner does not download models or contact a cloud inference endpoint. From this extracted directory: python3 scripts/evaluation/run.py --model qwen2:0.5b --output data/evaluation/document-extraction.json python3 scripts/evaluation/run.py --model llama3:latest --output data/evaluation/document-extraction-llama3.json python3 scripts/evaluation/publish.py Start your own local Ollama server before running. Model versions, hardware, caches and software versions can change results. Original model digests, prompts, options, fixture hash, raw outputs, failures and latency are in the recorded JSON files. Recorded latency is one local wall-time measurement per case after warm-up. It is not a cloud cost or a client saving. The second model was chosen after inspecting the first run. Do not represent this exercise as a pre-registered or representative benchmark. Task correctness includes the expected review decision. Correct abstention is not an automated successful extraction. All observed failures remain in the package. Ollama generate API: https://docs.ollama.com/api/generate Public explanation: https://coreline.agency/resources/document-extraction-evaluation/