Grounding
Can the system support every claim with authorized, current evidence?
Open research
Evaluate AI against regional reality—not a translation of it.
We build native benchmarks for tasks where context changes the answer: freight operations, healthcare purchasing, financial risk, and customer support. Every case begins with authorized regional evidence and professional review.
Explore the research ↓01Local evidenceDocuments, records, and rules from the market being evaluated
02Real tasksComplete decisions, not isolated knowledge questions
03Cost-aware errorsFailures weighted by their operational consequence
Preprint releaseSC—CO / v0.1.0
es-CO · FREIGHT OPERATIONS
Savia Carga · Colombia
An evaluation program for operational reasoning in road freight: local terminology, grounded calculations, exception handling, calibration, and safe abstention.
Source-groundedExpert reviewedes-CO nativeReproducible
Version 0.1.0 is published with its answer key, derivations, baseline result and sources. The questions have not yet been confirmed by a domain expert, and the page says so. Open the release ↗
What we measure
01
Source fidelity
Does the answer stay inside the evidence?
02
Operational correctness
Would the decision work in practice?
03
Calibrated abstention
Does the system know when not to answer?
Research agenda
Questions that matter in production.
Regional judgment
Does it recognize when a rule, term, or practice changes across countries?
Safe abstention
Does it stop and escalate when the evidence cannot support a responsible answer?
Methodology
A benchmark should explain why a system fails, not only how often.
Each release documents authorization, case origins, expert criteria, regional assumptions, and use limits. Results include error patterns and the evidence required to reproduce them.
- 01Regional task definition
- 02Authorized evidence set
- 03Expert scoring rubric
- 04Transparent error analysis