Scarab Systems Field Lab: Same Patch, 96% Less Final Decision Context

Short summary

Scarab Systems benchmarked diagnostic governance (SDS) against a baseline agent on a real code-repair task. Both produced the same one-line fix, but SDS used 96% fewer tokens in decision context (2,651 vs 70,212). The insight: AI agents don't need bigger context windows—they need smarter governance layers to surface the narrow truth boundaries they actually need.

• •Diagnostic governance layer reduced final decision context by 96% while maintaining identical repair quality
• •Challenges the architectural assumption that AI agents require large context payloads for accuracy
• •Proposes ephemeral, task-specific governance over permanent agent memory—repo truth lives in governance, not in the agent

Generated with AI, which can make mistakes.