GeoAnalystBench · Task 34 · 2026-07-27

One real task.
Three separate claims.

Estimate the accessibility of roads to rural areas in Japan. This record separates permission to use the task, fidelity to published benchmark gold, and what the agents actually achieved.

OPERATING PATH PROVED

Evidence claims

do not collapse these into one score

1 · Rights scope

Academic research and dissemination

Repository owner attested academic research and dissemination use on 2026-07-27.

Input handling: private inputs; not uploaded to Arena. Only generated results, maps, code, and hashes are published.

Upstream benchmark source

2 · Benchmark gold

Reproducible image, incomplete semantics

The bundled reference image reproduced pixel-for-pixel: yes.

The reference code maps an area fraction but does not compute the requested aggregate population percentage. Benchmark match and spatial validity therefore require separate verdicts.

3 · Agent performance

8 model routes attempted

Two Kimi routes produced the benchmark-compatible numeric result and inspectable artifacts. Other routes produced alternate semantics, wrong selections, runtime errors, serialization failure, or no visible program.

Cloudflare Workers AI · local bubblewrap sandbox with network namespace isolated

The answer depends on which claim you test

Published tutorial

736,316 / 872,986

about 84 percent

Benchmark-compatible · EPSG:4087

84.3349%

736231.6 estimated accessible residents.

Spatial-validity audit · EPSG:6672

85.9759%

750557.2 estimated accessible residents. This remains an audit result, not accepted benchmark gold.

Choropleth map generated for GeoAnalystBench Task 34 showing estimated rural road-access fractions across Shikoku
Generated evidence artifact from moonshotai/kimi-k2.6; source datasets are not served by the Arena.

Observed model outcomes

single routed observations, not reliability estimates
Actual generation and isolated execution outcomes for the eight Workers AI lanes.
ModelAttemptsTerminal stateBenchmark matchSpatial validityReported accessFailure detail
moonshotai/kimi-k2.62executedMATCHNOT VALIDATED84.3349%
zai-org/glm-4.7-flash1runtime_errorNO RESULTNO RESULTGeoSeries has no dissolve method
zai-org/glm-5.21executedNO MATCHNOT VALIDATED86.2101%
moonshotai/kimi-k2.7-code1executedMATCHNOT VALIDATED84.3349%
openai/gpt-oss-20b1executed_wrong_resultNO RESULTNO RESULT78.9400%
google/gemma-4-26b-a4b-it2no_visible_programNO RESULTNO RESULToutput limit exhausted
qwen/qwen3-30b-a3b-fp82runtime_errorNO RESULTNO RESULTNumPy int64 was not JSON serializable
mistralai/mistral-small-3.1-24b-instruct1runtime_errorNO RESULTNO RESULTmunicipality layer has no ID field

Inspectable artifacts

machine-readable evidence

What this does not prove