nawnie/atlas-lora-adapter ? reverse-engineered prompt

Build me an early research lab for testing whether a small LoRA adapter can teach a language model to read structured retrieval context better instead of treating it like a big text dump.

I want the project to use Atlas style lanes and cards, with compact evidence packs, clear source priority, boundary rules, and a fallback when the answer is not supported by the Atlas evidence. It should support training and evaluation for this behavior, plus simple scripts to validate the data and summarize results. Please keep it focused on the research workflow, not a production app.

Make sure the repo is easy to navigate, with docs for the strategy, the lane system, the training data, and the evaluation results. Include a way to run the main checks locally, and make the setup clear enough that someone can reproduce the current experiments. If you need to, look up current docs online for the model and training stack.