Agent regression testing benchmark: embeddings caught only 7/19 real answer changes

redhatpanda07 · reddit · 2026-09-23

The author built a 109-pair fixture with ground truth by construction (19 real changes, 90 rewordings) to benchmark detectors for agent prompt-edit regression. Exact match: 19/19 caught but 84/90 false alarms; embeddings under 0.75 caught only 7/19; a markers→embed→LLM-judge stack got 19/19 with 33 false alarms; a 4-way Choice classifier hit 19/19 with 12 flags at $0.0028 and 0.43s. Key lesson: false alarms, not recall, are the metric — every false alarm costs a human re-read.

Original post →

More from coding & agent

coding & agent channel →