All of Taylor Amarel’s public AI safety benchmarks, OpenAI Red-Teaming Honorable Mention notebooks, DueCare 6-lane harness-lift studies (55+ notebooks), Jev & JED/JC GPU T4 recognition probes, and Philippine migrant-worker jurisprudence datasets in one place.
| Track | Notebook / Dataset Title & Direct Kaggle Link | What It Tests, Proves, or Provides |
|---|---|---|
| Public Benchmark | Dataset: duecare-jev-exploitation-safety-benchmark | Unified public Kaggle dataset packaging Jev (jev-1.13.0) vs. Rules vs. Gemma 4, DueCare 6-lane, and JC T4 probe tables. |
| Public Benchmark | Notebook: duecare-jev-human-exploitation-safety-benchmark | Reproducible starter notebook loading and scoring the unified Jev, DueCare, and Recognition-Gated Safety benchmark. |
| OpenAI Red-Team | llms-support-human-trafficking | Primary study for our 2025 OpenAI Red-Teaming Challenge Honorable Mention uncovering 5 classes of debt-bondage complicity. |
| OpenAI Red-Team | v4-red-teaming-challenge-starter-notebook (178 Upvotes) | Community-adopted red-teaming evaluation harness (part of 760+ upvotes across v1–v4 starter notebooks). |
| OpenAI Red-Team | red-teaming-challenge-starter-notebook (v1) (106 Upvotes) · v2 (95) · v3 (65) | Complete progression of open red-teaming harnesses for probing `gpt-oss-20b` on Kaggle GPUs. |
| OpenAI Red-Team | native-business-framed-exploitation-set-1 | Probes how routine HR, payroll, and agency framing induces models to draft coercive fee-shifting clauses. |
| OpenAI Red-Team | jurisdictional-hierarchy-exploitation-set-1 | Tests cross-border conflict-of-laws blindness (Philippines RA 10022 vs. Hong Kong Cap. 57/163 vs. ILO C181). |
| OpenAI Red-Team | legal-standards-financial-crime-blindness-set-1 | Evaluates 68% APR usury novation and cross-border money-lending compliance failures. |
| OpenAI Red-Team | migrant-worker-re-victimization-set-1 | Demonstrates how unassisted models validate trafficker framing and reinforce guilt when a victim asks for help. |
| OpenAI Red-Team | amplification-through-known-attacks-set-1 · gemma-4-e4b-red-teaming-human-exp-w-tools | Multi-turn amplification and tool-assisted exploitation red-teaming on `gpt-oss-20b` and `Gemma 4 E4B`. |
| JC / JED Probe | jc-realmodel-t4 (GPU T4 Real-Model Probe) | 16-scenario evaluation on Qwen2.5-7B-Instruct showing 93.8% out-of-band recognition vs. 81.3% framed compliance. |
| JC / JED Probe | jc-strong-t4 (12-Scenario Strong Subset on GPU T4) | Demonstrates 91.7% framed compliance dropping to 8.3% (and 83.3% → 0.0% on conflict-of-laws) with a recognition gate. |
| JC / JED Probe | jed-attack-submission | Multi-step Format-as-Affordance agent tool hijacking (EXFILTRATION & CONFUSED_DEPUTY) benchmark. |
| DueCare 6-Lane | duecare-harness-lift-benchmark-start-here | Entry point to the 55-notebook DueCare multi-stakeholder safety evaluation suite. |
| DueCare 6-Lane | duecare-what-the-benchmark-proves · duecare-does-a-safety-harness-help | Empirical analysis of harness lift across exploitative vs. matched employer-paid benign twins. |
| DueCare 6-Lane | duecare-four-arm-fine-tuning-evaluation · duecare-gemma4-four-arm-before-after | Compares Stock Gemma 4, Suspicion Prompt, Unfiltered RAG (negative result), and the 6-Layer DueCare Harness. |
| DueCare 6-Lane | duecare-judge-panel-calibration · duecare-deterministic-verification · duecare-rulecard-witness-atlas | Deterministic statutory clause verification, rulecard witness tracing, and multi-judge calibration. |
| PH Law & Data | dmw-job-orders-v2-wordpress-post-version (18 Upvotes) · eda-philippines-dmw-job-order-dataset | Exploratory analysis of official Philippine Department of Migrant Workers (DMW) approved job orders across corridors. |
| PH Law & Data | sec-express-monitoring-script · philippines-jurisprudence-database-analysis | Corporate registry monitoring and Supreme Court anti-trafficking / illegal-recruitment jurisprudence analysis. |