Site & constraint reasoning
Does the answer account for overlapping constraints and avoid treating screening evidence as a final suitability decision?
GeoBench by Nexora AI
The benchmark for professional geospatial reasoning.
Evaluate how models interpret spatial evidence, reason through physical constraints, and support professional decisions. Inspect the task, the response, and the criteria behind each result.
Spatial
Infrastructure
Environment
Energy
GeoBench / Task viewer
Inspect the evidence. Evaluate the answer.
Illustrative scores and example tasks; not measured model results.
TASK 01 / 03
GB-001Site reasoning · Evaluation task
Site Suitability Analysis
Two port-expansion sites intersect a mapped flood zone. Which site should advance to further diligence?
Supplied evidence
- Site A: 18% flood-zone overlap; road access confirmed.
- Site B: 6% flood-zone overlap; road access unknown.
- No engineering, permitting, or flood-depth evidence is supplied.
Reference expectation
Do not select a definitive winner. Site B has less mapped overlap, but its access is unverified. Compare both constraints and identify the missing diligence.
Model response (illustrative)
Site B has less mapped flood-zone overlap (6% versus 18%). It may warrant further review, but road access needs verification. These layers alone cannot establish development suitability.
Evaluation findings
The response grounds its comparison in both overlap values and avoids a definitive suitability claim. It could more explicitly retain Site A as a candidate pending further evidence.
Evaluation rubric
Illustrative task scores · 0–100 per criterion
90.0 / 100
Illustrative weighted score
- Spatial accuracy 40%weight90/100
- Correct relationships, measurements, and interpretation of the supplied spatial evidence.
- Evidence grounding 30%weight95/100
- Claims trace back to the supplied sources without inventing facts or capabilities.
- Reasoning quality 20%weight80/100
- The conclusion follows from the evidence and accounts for relevant constraints.
- Uncertainty 10%weight95/100
- Missing evidence and limitations are stated; unsupported conclusions are withheld.
Weighted score = sum of each criterion score × its weight. Acceptance criteria and thresholds are defined for the use case.
Compare the dimensions
Compare model performance.
Understand where each model performs well and where closer review is needed. Compare evaluation dimensions or sort by the weighted score.
Model performance
| Model | |||||
|---|---|---|---|---|---|
| Model 1Evidence-grounded workflow | 88.0 | 92.0 | 84.0 | 90.0 | 88.6 |
| Model 2General-purpose baseline | 76.0 | 71.0 | 80.0 | 65.0 | 74.2 |
| Model 3Spatial tool-assisted workflow | 93.0 | 86.0 | 89.0 | 82.0 | 89.0 |
Illustrative scores and example tasks; not measured model results. Higher is better within this rubric.
Model order: 1, 2, 3.
Transparent criteria
Make the scoring inspectable.
Assess spatial accuracy, evidence grounding, reasoning quality, and uncertainty using explicit criteria. Weight each dimension around the decision and its requirements.
Evaluation rubric
Scoring methodology · weights total 100%
- Spatial accuracy 40%weight
- Correct relationships, measurements, and interpretation of the supplied spatial evidence.
- Evidence grounding 30%weight
- Claims trace back to the supplied sources without inventing facts or capabilities.
- Reasoning quality 20%weight
- The conclusion follows from the evidence and accounts for relevant constraints.
- Uncertainty 10%weight
- Missing evidence and limitations are stated; unsupported conclusions are withheld.
Weighted score = sum of each criterion score × its weight. Acceptance criteria and thresholds are defined for the use case.
Benchmark coverage
Tasks tied to the physical world.
Evaluate the skills that connect geographic evidence to professional decisions: interpreting constraints, understanding networks, and explaining change.
Networks & accessibility
Does the system reason over valid connections, route restrictions, and the difference between proximity and reachability?
Change & interpretation
Can it separate observable change from inferred causes, and identify what evidence is missing?
From evaluation to action
Turn performance into progress.
Use evaluation findings to prioritize model improvements and build more dependable workflows.
Find the failure pattern
Separate errors in spatial interpretation, source use, reasoning, and uncertainty. Identify the conditions that lead to unreliable answers.
Choose the next improvement
Use task-level findings to refine retrieval, tool use, instructions, and review requirements. Focus effort on the gaps that matter to the intended workflow.
Track changes consistently
Compare model and workflow revisions against the same versioned tasks and criteria. Review both aggregate performance and the cases behind it.
Nexora AI
Start with a clearly defined need.
Tell us about your project. Scope and scheduling are confirmed after review.
A focused project. A practical path forward.