Selection and validation
These sets prioritize difficult examples while retaining scenario coverage. Supplements mean membership is not equivalent to one universal n-success threshold. They are test challenge sets; do not train on them and then present evaluation as held out.
| Configuration | Validation and selection notes |
|---|---|
| text_adversarial_curated | 141 user human-review passes after earlier checks: 129 GPT-5 semantic confirmations, 2 earlier user adjudications, 10 user-authorized Sensor inclusions outside that GPT-5 pass. All 141 received the subsequent human review. |
| text_injection_curated | 374 retained under the explicit instruction-hijacking/original-gold protocol. No uniform external dual semantic validation. Failure-based selection with source balancing. |
| image_perturbation_curated | 200 dual-label-verified candidates: 127 common perturbations and 73 original-size ε=8/255 optimization candidates. Human image-consistent exclusions removed 87 flagged scoring items plus 4 counterparts sharing the same PNG. Prior retained items are not relabelled explicit readability passes. Version 1.2.0 adds 13 n=1 candidates from the original five severity levels, all dual-verified and human-approved (unmarked explicitly accepted by the user), restoring six originals. These additions have at most three distinct image variants per original, one scoring mode per image; ties in n prioritize probability-margin evidence. No low-strength exploratory variants are included. |
| image_overlay_curated | 223 candidates: 204 actual dual-model correct/readable passes and 19 approved Claude-only exceptions after Azure GPT-5 filtering. At most three candidates per original; includes two n=0 spatial-counting supplements. |
| image_instruction_curated | 120 candidates: 117 with n≥2 and three n=1 scene supplements. No external validation pass. |
Here n is the number of selection models successfully attacked (clean correct and adv wrong), not merely the number giving a wrong adversarial answer. Semantic preservation and visibility checks can still contain errors; the checks above are not all identical. The release preserves original gold labels, full requests and native scoring modes.