GIS-MAS: derived spatial data, query sets, evaluation artefacts and code for a spatial-reasoning-enhanced collaborative multi-agent framework

## Contents - `gis_mas/`, `experiments/`, `core/`, `operators/`, `run_*.py`, `config.py`, `requirements.txt`, `.env.example` - the implementation and every evaluation script cited in the paper. The entry point for the full evaluation is `experiments/eval/run_all.py`; the end-to-end truth tier is module O and must be invoked explicitly. - `gis_mas/data/geo_knowledge.json` - the 672-entry knowledge base (403 entities + 216 spatial relations + 53 operator documents), per-entry sources retained. - `data/*.geojson`, `data/rcc8_composition.json`, `data/reference/rcc8/`, `data/templates/*.yaml` - derived spatial layers (ODbL 1.0, (c) OpenStreetMap contributors, attribution retained), the audited RCC-8 composition table with its provenance block, and the 19 task templates. - `data/*.json` - the query sets: 227 evaluation queries, the 102-item synthetic benchmark, the 24 service-area expansion queries, and the held-out slice definition. - `output/eval_results/`, `output/ablation/` - the per-item evaluation artefacts cited in the paper (the evidence record), plus the blind annotation sheets and the machine/human annotation readings. - `MANIFEST_SHA256.txt` - SHA256 and size of every file in this deposit. ## Deliberately NOT included - **Amap-derived data** (POI snapshot, geocoding results, driving-distance truth): the Amap Open Platform terms of service do not permit redistribution. These files are reproducible only by re-collecting with the reader's own API key; the statistical scope and grid parameters are given in the paper (Table 1). - **OpenStreetMap extracts** (`.gpkg`, `.osm.pbf`) and derived road-network graphs (`.pkl`): hundreds of MB, ODbL-encumbered, and rebuildable with the scripts in `experiments/build_*.py`. - **Model caches and third-party binaries** (embedding models, GraphHopper jar) and **build products** (rendered PNGs, logs). ## Licences This deposit is released under **CC BY 4.0** (the record licence), with two stated exceptions that keep upstream terms intact: - **Code** (`gis_mas/`, `experiments/`, `run_*.py`, `core/`, `operators/`): **MIT** (see `LICENSE`). - **OpenStreetMap-derived spatial layers** (`data/*.geojson`): **ODbL 1.0**, (c) OpenStreetMap contributors. - Wikidata-derived attributes: CC0. ## Reproduction Install `requirements.txt` in a Python 3.13 environment, copy `.env.example` to `.env` and add your own Amap and DashScope keys, then run `python experiments/eval/run_all.py` (about 40-60 minutes; module O runs separately). Service-area and network-distance steps need GraphHopper and a JDK; without them those steps degrade explicitly and say so in the answer. `output/eval_results/data_hashes.json` records the SHA256 of the shipped artefacts and can be regenerated with `python experiments/data_hashes.py`.