Structured Compare Calibration Summary

• generatedAt: 2026-03-22T10:44:18.102Z
• outputRoot: D:\Dev\myProject\prompt-optimizer\docs\workspace\compare-evaluation-analysis\structured-compare-calibration\latest

Case	Kind	Score	targetVsBaseline	targetVsReferenceGap	stopRecommendation	Expectation Match
live-basic-system-boundary-control	live	75	improved	minor	continue	exploratory
synthetic-medical-latent-trigger-overfit	synthetic	35	regressed	major	review	3/5
synthetic-ecommerce-schema-no-model-worship	synthetic	40	regressed	minor	review	6/6
synthetic-legal-flat-not-unclear	synthetic	50	flat	none	continue	3/3
synthetic-teaching-overfit-regression	synthetic	30	regressed	major	review	6/6
synthetic-hiring-replica-semantic-instability	synthetic	65	improved	none	review	4/4

Notes

• synthetic cases 用来检验 judge / synthesis 的提示词边界。
• live case 用来观察真实 target/teacher 执行结果在 structured compare 下是否能收敛成合理结论。
• 每个 case 子目录内都保存了 compare request、compare result、rewrite input / output，以及完整 LLM 调用日志。