Benchmarking an Agent Harness With the Model Held Fixed
A leaderboard score belongs to a model and a harness together. Here's how pinning the model isolates the harness effect, and where that breaks down.

A leaderboard score belongs to a model and a harness together. Here's how pinning the model isolates the harness effect, and where that breaks down.

Multi-agent orchestration in the Geocluster Research Harness: each specialist runs as its own child process, with credentials kept out of logged commands.

MCP tool design lessons from real agents: they ignored structuredContent, invented legend labels and misread 86 records as "1 item". Here's what we changed.

Geological map legend extraction in Stratigraphic Amenity: a YOLOv10 detector finds the map parts, then swatches are paired to text and colours read.

Geochemistry data cleaning for AI agents: why "<0.005" and comma decimals break code, and how our harness diagnoses data before cleaning, without guessing.

Georeferencing scanned maps with an affine fit from ground control points, and why a leave-one-out check catches a misread coordinate that least squares hides.

Agents with file and shell tools can reach the grading rubric on disk. What we measured across 50 questions, which fixes failed, and what the harness does now.

Learn how to build a new world model via coherence mapping by connecting geological observations, reducing uncertainty, and finding relationships that explain mineral systems.

Best AI for Geology Testing. See how Groundtruth's dynamic benchmark tests models on your own data, comparing accuracy, cost, and performance.

Learn five harness design principles to avoid building a slop cannon, plus a real-world mining industry case study on improving AI agent output.