supercarto-bench

Benchmark fixtures, harness results, and raw runs comparing map-data representations for LLM agents: a compiled topological graph (supercarto maplets) vs raw GeoJSON vs no map, scored on answer accuracy and context-token cost.

Fixtures are unmemorizable synthetic micro-areas plus eight Overpass-derived city extracts. Raw run transcripts (JSONL) and per-run report cards come from the released supercarto npm package; the harness, and the full history, lives at https://github.com/Unselfisheologism/supercarto-bench — the mirror here is read-only convenience for HF users.

Contents

Path	What it is
fixtures/bench/	Synthetic and city GeoJSON fixture areas
fixtures/overpass/	Cached Overpass API extracts (.json.gz)
bench-results/*.jsonl	Raw question/answer transcripts per run
bench-results/*.report.json	Scored summaries per run
bench-results/leaderboard.json, leaderboard.md	Current headline results
bench-results/certificate.*	Full-suite pass certificate
docs/datasheet.md	Full datasheet (motivation, composition, collection)
docs/benchmark.md, docs/benchmark-sop.txt	Method and the SOP it was worked from
croissant.json	Croissant metadata (also machine-readable on GitHub)

Headline (both official free models)

	supercarto	geojson	none
omission / nearest / connectivity	9/9	~4/9	0/9
expansion (footprint m²)	5/6	3/9	0/9
heat (peak value + location)	6/6	0/6	0/6
diff (previous map + changes)	5/5 answered	4/4	0/6

License

Code and harness: Apache-2.0. GeoJSON fixtures derived from OpenStreetMap: ODbL (attribution in fixtures/bench/ATTRIBUTION.md and fixtures/overpass/).