Merge pull request #93 from wuyoscar/update/2026-04-24

update/2026-04-24: template MIN_WORDS + TVD diversification + aiml_guard rename

Author: wuyoscar

AGENT_README.md

@@ -25,7 +25,7 @@ We study this with the **TVD** (Task + Validator + Data) framework:
 - **Validator**: tool validation that defines whether the task succeeds
 - **Data**: sensitive data fields the model must fill in to pass validation
 
-**77 templates** across **9 domains** (AI/ML, biology, chemistry, cybersecurity, epidemiology, clinical genomics, pharmacology, media, and more). **28 confirmed models** including GPT-5.2, Claude Opus 4.6, Gemini 3 Pro, GLM-4.7, and others.
+**84 templates** across **9 domains** (AI/ML, biology, chemistry, cybersecurity, epidemiology, clinical genomics, pharmacology, media, and more), with a growing ISC Arena of confirmed and ranked model reproductions.
 
 ### Boundaries
 

CHANGELOG.md

@@ -75,7 +75,7 @@ ISC can still be triggered on live production endpoints as of 2026-04-10. Attack
 ### Community Evidence (10 new folders)
 - Multi-template evidence for Gemini 3 Flash (guard_attack_v2 + detoxify + offensive)
 - Multi-template evidence for GPT-5.1 (guard_attack_v2 + detoxify + phishing)
-- Multi-template evidence for Grok 4.1 (guard_attack_v2 + detoxify + aiml_guard)
+- Multi-template evidence for Grok 4.1 (guard_attack_v2 + detoxify + aiml_qwen3guard)
 
 ### Agent Fix
 - Fix `reasoning_effort` bug in `experiment/isc_agent/agent.py`: was applied to all models; now only set when `--thinking` flag is passed

ISC_PAPER_DIGEST.md

@@ -448,7 +448,7 @@ The key variation across domains is in the **validator** and **anchor**:
 
 For single-turn: everything in one prompt file. For agentic mode: split into separate files in a workspace directory.
 
-All 53 templates are in [`templates/`](templates/) — browse them for more domain-specific patterns.
+All 84 templates are in [`templates/`](templates/) — browse them for more domain-specific patterns.
 
 ## TVD Customization
 

README.md

@@ -466,7 +466,7 @@ Quickest way in: open any template folder. `prompt.txt` is the evaluation prompt
 
 | # | Task | Tool | Template |
 |:-:|------|------|:--------:|
-| 46 | LLM input-output safeguard classification | Llama-Guard | [📄](templates/aiml_guard/) |
+| 46 | LLM input-output safeguard classification | Llama-Guard | [📄](templates/aiml_llamaguard/) |
 | 47 | Statistical outlier detection | PyOD | [📄](templates/aiml_pyod/) |
 | 48 | Toxic comment classification | Detoxify | [📄](templates/aiml_detoxify/) |
 | 49 | Content moderation benchmark (3 variants) | OpenAI Moderation API | [📄](templates/aiml_moderation/) |
@@ -495,7 +495,7 @@ Quickest way in: open any template folder. `prompt.txt` is the evaluation prompt
 </details>
 
 ```bash
-cat templates/aiml_guard/prompt.txt
+cat templates/aiml_llamaguard/prompt.txt
 # → Copy, paste into any LLM. That's it.
 ```
 

README_es.md

@@ -464,7 +464,7 @@ La vía más rápida: abre cualquier carpeta de plantilla. `prompt.txt` es el pr
 
 | # | Tarea | Herramienta | Plantilla |
 |:-:|------|------|:--------:|
-| 46 | Clasificación de salvaguardas de entrada/salida de LLMs | Llama-Guard | [📄](templates/aiml_guard/) |
+| 46 | Clasificación de salvaguardas de entrada/salida de LLMs | Llama-Guard | [📄](templates/aiml_llamaguard/) |
 | 47 | Detección estadística de valores atípicos | PyOD | [📄](templates/aiml_pyod/) |
 | 48 | Clasificación de comentarios tóxicos | Detoxify | [📄](templates/aiml_detoxify/) |
 | 49 | Benchmark de moderación de contenido (3 variantes) | OpenAI Moderation API | [📄](templates/aiml_moderation/) |
@@ -493,7 +493,7 @@ La vía más rápida: abre cualquier carpeta de plantilla. `prompt.txt` es el pr
 </details>
 
 ```bash
-cat templates/aiml_guard/prompt.txt
+cat templates/aiml_llamaguard/prompt.txt
 # → Copia y pega en cualquier LLM. Eso es todo.
 ```
 

README_ja.md

@@ -464,7 +464,7 @@ ISC はフロンティア大規模モデル上で現れ続けています。以
 
 | # | タスク | ツール | テンプレート |
 |:-:|------|------|:--------:|
-| 46 | LLM 入出力セーフガード分類 | Llama-Guard | [📄](templates/aiml_guard/) |
+| 46 | LLM 入出力セーフガード分類 | Llama-Guard | [📄](templates/aiml_llamaguard/) |
 | 47 | 統計的外れ値検出 | PyOD | [📄](templates/aiml_pyod/) |
 | 48 | 有害コメント分類 | Detoxify | [📄](templates/aiml_detoxify/) |
 | 49 | コンテンツモデレーションベンチマーク（3バリアント） | OpenAI Moderation API | [📄](templates/aiml_moderation/) |
@@ -493,7 +493,7 @@ ISC はフロンティア大規模モデル上で現れ続けています。以
 </details>
 
 ```bash
-cat templates/aiml_guard/prompt.txt
+cat templates/aiml_llamaguard/prompt.txt
 # → コピーして任意の LLM に貼り付けるだけです。
 ```
 

README_ko.md

@@ -463,7 +463,7 @@ ISC는 프론티어 대형 모델에서 반복적으로 나타납니다. 아래
 
 | # | 작업 | 도구 | 템플릿 |
 |:-:|------|------|:--------:|
-| 46 | LLM 입출력 보호 장치 분류 | Llama-Guard | [📄](templates/aiml_guard/) |
+| 46 | LLM 입출력 보호 장치 분류 | Llama-Guard | [📄](templates/aiml_llamaguard/) |
 | 47 | 통계적 이상치 탐지 | PyOD | [📄](templates/aiml_pyod/) |
 | 48 | 독성 댓글 분류 | Detoxify | [📄](templates/aiml_detoxify/) |
 | 49 | 콘텐츠 조절 벤치마크(3가지 변형) | OpenAI Moderation API | [📄](templates/aiml_moderation/) |
@@ -492,7 +492,7 @@ ISC는 프론티어 대형 모델에서 반복적으로 나타납니다. 아래
 </details>
 
 ```bash
-cat templates/aiml_guard/prompt.txt
+cat templates/aiml_llamaguard/prompt.txt
 # → 복사하여 어떤 LLM에든 붙여넣으세요. 그것으로 끝입니다.
 ```
 

README_pt.md

@@ -463,7 +463,7 @@ Forma mais rápida de começar: abra qualquer pasta de template. `prompt.txt` é
 
 | # | Tarefa | Ferramenta | Template |
 |:-:|------|------|:--------:|
-| 46 | Classificação de salvaguarda de entrada-saída de LLM | Llama-Guard | [📄](templates/aiml_guard/) |
+| 46 | Classificação de salvaguarda de entrada-saída de LLM | Llama-Guard | [📄](templates/aiml_llamaguard/) |
 | 47 | Detecção estatística de outliers | PyOD | [📄](templates/aiml_pyod/) |
 | 48 | Classificação de comentários tóxicos | Detoxify | [📄](templates/aiml_detoxify/) |
 | 49 | Benchmark de moderação de conteúdo (3 variantes) | OpenAI Moderation API | [📄](templates/aiml_moderation/) |
@@ -492,7 +492,7 @@ Forma mais rápida de começar: abra qualquer pasta de template. `prompt.txt` é
 </details>
 
 ```bash
-cat templates/aiml_guard/prompt.txt
+cat templates/aiml_llamaguard/prompt.txt
 # → Copie, cole em qualquer LLM. Só isso.
 ```
 

README_vi.md

@@ -463,7 +463,7 @@ Cách vào nhanh nhất: mở bất kỳ thư mục template nào. `prompt.txt`
 
 | # | Tác vụ | Công cụ | Template |
 |:-:|------|------|:--------:|
-| 46 | Phân loại bảo vệ input-output LLM | Llama-Guard | [📄](templates/aiml_guard/) |
+| 46 | Phân loại bảo vệ input-output LLM | Llama-Guard | [📄](templates/aiml_llamaguard/) |
 | 47 | Phát hiện ngoại lệ thống kê | PyOD | [📄](templates/aiml_pyod/) |
 | 48 | Phân loại bình luận độc hại | Detoxify | [📄](templates/aiml_detoxify/) |
 | 49 | Benchmark kiểm duyệt nội dung (3 biến thể) | OpenAI Moderation API | [📄](templates/aiml_moderation/) |
@@ -492,7 +492,7 @@ Cách vào nhanh nhất: mở bất kỳ thư mục template nào. `prompt.txt`
 </details>
 
 ```bash
-cat templates/aiml_guard/prompt.txt
+cat templates/aiml_llamaguard/prompt.txt
 # → Sao chép, dán vào bất kỳ LLM nào. Vậy là xong.
 ```
 

README_zh.md

@@ -473,7 +473,7 @@ ISC 在前沿大模型上持续出现。下面这些案例由社区贡献者完
 
 | # | Task | Tool | Template |
 |:-:|------|------|:--------:|
-| 46 | LLM input-output safeguard classification | Llama-Guard | [📄](templates/aiml_guard/) |
+| 46 | LLM input-output safeguard classification | Llama-Guard | [📄](templates/aiml_llamaguard/) |
 | 47 | Statistical outlier detection | PyOD | [📄](templates/aiml_pyod/) |
 | 48 | Toxic comment classification | Detoxify | [📄](templates/aiml_detoxify/) |
 | 49 | Content moderation benchmark (3 variants) | OpenAI Moderation API | [📄](templates/aiml_moderation/) |
@@ -502,7 +502,7 @@ ISC 在前沿大模型上持续出现。下面这些案例由社区贡献者完
 </details>
 
 ```bash
-cat templates/aiml_guard/prompt.txt
+cat templates/aiml_llamaguard/prompt.txt
 # → Copy, paste into any LLM. That's it.
 ```