JUST-OS ideas

- Make top-k adaptive: over-retrieve, rerank, then apply a relative cutoff (keep within X% of the top score) bounded by a min and max k. - so that targeted questions don't get junk in, and broad questions get sufficient coverage
- Benchmark against a more modern 8B model, given that we don't use the full OpenScholar pipeline (e.g. Gemma 4 E4B, which scores 'highest' on AA Omniscience - i.e. low hallucinations)
- Benchmark against the actual OpenScholar pipeline (generation-feedback-regeneration)
- Curate our own benchmark, with expert-written OS answers (not sure how much of that exists already)