Hi organisers and fellow participants — 4 clarifications that will affect many solutions. Thanks in advance!

1. How strictly is the 3B parameter cap counted?

The rules say one generative model "with no more than 3 billion parameters." Many models marketed as "3B" actually exceed 3.0B raw parameters (e.g. SmolLM3-3B = 3.08B, Qwen2.5-3B = 3.09B, Llama 3.2 3B = 3.21B). Does the cap mean (a) the nominal 3B class is acceptable, or (b) a literal hard ceiling of 3.0B? And for architectures where "effective" and "total" parameter counts differ (e.g. Gemma's E2B/E4B per-layer-embedding design), which count applies?

2. Are decoder-backbone embedding/reranker models permitted tools?

The rules permit "tools for retrieval, OCR, embeddings, reranking..., provided they do not invoke additional generative models." Some embedding/reranking models are built on decoder-LLM backbones (e.g. Qwen3-Embedding-0.6B, Qwen3-Reranker-0.6B) but are used purely to output fixed vectors or relevance logits — they never generate text or answers. Do these qualify as permitted "embeddings"/"rerankers" tools despite being architecturally decoder-based?

3. Is there a required citation/evidence format?

For the "evidence identifying the original source pages, cells, records, features or model elements": is a specific format required (e.g. SourceID.pdf#page=12, or a structured field per Test ID), or is free-form natural-language referencing acceptable as long as it is specific and verifiable? Would a structurally valid answer whose citation is format-mismatched but findable lose evidence credit?

4. Do prebuilt index artifacts count as "custom packages"?

The reproducibility rules say "custom packages in your submission notebook will not be accepted." May the submission repo include data artifacts built offline from the provided source materials (OCR'd text, chunked text, embedding indexes, parsed 3DBAG JSON) as long as scripts to regenerate them from the originals are included and no additional generative models were involved?

Happy to split these into separate threads if preferred.