How Four AI Projects Used Code Checks to Catch Model Errors Before They Caused Harm

Four open-source projects—Gilbeot, Sentinel, AirBridge, and Project Rosie—each independently adopted a strategy of validating AI model outputs through deterministic code checks rather than relying solely on prompts. Gilbeot, a walking assistant for elderly users in Korea, uses coordinate comparisons to override left-right direction errors from a small vision model. Sentinel, a security scanning tool, rejects GPT-generated reviews that reference unseen lines or invalid IDs, retrying until a valid response passes schema checks. AirBridge enforces a tool catalog with strict argument validation, returning refusals as structured feedback to the model rather than raising exceptions. Project Rosie replaced AI-generated synthesis documents with fixed templates after recognizing that any fabricated values—like catalog numbers or QC thresholds—could undermine clinical trust, reserving the model only for prose-based tasks.