Project 03 — Educational baseline v1

Project 03 educational baseline: a Llama 3.1 8B QLoRA adapter experiment with documented training and a matched comparison against gpt-4.1-2025-04-14.

The attached evidence package includes frozen inputs, saved outputs, reviews, training records, licenses, and a CPU replay that checks artifact hashes and reproduces scores. Extract the archive and run python3 educational-v1/replay.py. This replays saved evidence; it does not run live model inference.

On 200 balanced simulated classification cases, adapter F1 was 89.66% versus 63.80%. GPT-4.1 performed better on supported risk rationales and unfamiliar regulatory citations. Read the full comparison report for task-specific results, review limitations, and scoring sensitivity. No general GPT superiority is claimed. Original strict qualification remains BENCHMARK_GATES_FAILED.

The educational adapter is now hosted on Hugging Face. All 13 files in original revision 308021f3ba42940f09edbe787d0146d9343ac11a passed anonymous download and hash verification. The attached evidence archive does not contain adapter weights. The article, tutorial recording, and book chapter remain separate editorial deliverables.

Commit author and committer: Etherlabs-dev ethercess@gmail.com.