遇见数据集

Intent-Fidelity Testbed v1.2.0: IF-04 Reconfirmation Efficacy + IF-06 Adversarial Drift

收藏
Zenodo2026-08-14 更新2026-08-20 收录
官方服务:

资源简介:

FIDELITY series experiments IF-04 (Reconfirmation Efficacy) and IF-06 (Adversarial Drift), extending the intent-fidelity-testbed. Part of the Remnant Fieldworks ExecutionProof / CIF program. IF-04 Reconfirmation Efficacy: Tests whether the IFDV dominant axis correctly identifies ground-truth divergence type. Result: PASS — 99.16% attribution accuracy (118/119 true positives correctly attributed), recall 0.9917, zero false positives. Robustness across 10 seeds: [0.992, 0.998, 1.000]. IF-06 Adversarial Drift: Tests whether multi-axis sub-threshold adversarial transformations evade IFDV detection. Result: PASS — 8.33% evasion rate (10/120 evaded, 110 caught), well below 15% ceiling. Robustness across 10 seeds: [0.075, 0.093, 0.125]. Honest scope: synthetic corpus; NOT live-LLM study. See also: GitHub repository (pending push of IF-04/IF-06 runners).

提供机构:
Zenodo
创建时间:
2026-08-14
二维码
社区交流群
二维码
科研交流群
商业服务