Hi organisers, a few clarifications that change how solutions are built:

1. Source-specific tools. Many Test questions need native readers and the calculation conventions of each source family (GTFS routing, IFC spaces, LAZ grids, WorldCover pixel selection, etc.). Is it allowed to write deterministic readers and tools per source family, designed after inspecting the supplied Test sources and Test question wording, as long as no human answers individual questions and every answer is produced by the pipeline within the model limits? Is tuning prompts on Test question wording also allowed?

2. Offline preprocessing. If the same single generative model is used once before answering to read source materials (for example to transcribe the scanned Sanborn sheets or summarise a file schema), do those calls count against the 15-call / 12,000-token budget of each question that later uses the output, or must every generative call happen inside a question's budget?

3. Submission format limits. Is there a maximum file size for the Markdown submission, or a maximum length per answer or Evidence block? And may answers and evidence contain non-ASCII UTF-8 characters (ü, é, ², typographic quotes), which appear in several Test questions and sources?

4. Evidence for negative answers. When the correct answer is that the sources do not support a claim or the information is insufficient, what evidence earns full support credit? Should we cite the note, definition, field or missing value that shows the gap?

5. Reproduction run. Is there a wall-clock limit for regenerating the full 600-question submission on the RTX 4090, and will the container have internet access to download model weights and packages, or must they be bundled? Thank you!