Evaluate accuracy and speed.

Compare configurations, inspect missed detections and review pass or fail results.

Architecture

Bedrock proposes configuration changes using development results. OpenCV runs the evaluations. The selected configuration is frozen before separate test evaluation, where recall, precision and latency determine whether it passes.

Measured results and limits

A reduced-resolution PCB configuration lost 559 baseline detections and returned CI exit 2. Four workers preserved detections and passed. On 48 pedestrian test images, four workers preserved TP138/FP41/FN1 and recorded a 3.13x median paired speedup, with a 95% interval of 2.53-3.40x.

The 166 fresh PCB images come from three boards with earlier baseline calibration exposure. They are not unseen boards. Bedrock versus the fixed sweep had an interval including parity, and the sweep had better quality. We do not claim an LLM search advantage. AWS library builds differed, so the observed COOL comparison does not isolate COOL's effect.

Demo video

Download the demonstration · Source, reproduction and CI · Inspect the replay

Scope and attribution

This site shows recorded results and sample images. Metrics cover the full registered test sets. Fresh inference and evaluation run in the local app. Public exports omit private paths; hashes refer to the original evidence.

DeepPCB: Tang et al., MIT-licensed dataset. Penn-Fudan: Liming Wang, Jianbo Shi, Gang Song and I-fan Shen, research dataset. Images retain their original rights. YOLOX: OpenCV Zoo, Apache 2.0. Model weights and full datasets are downloaded separately.