A scenario-driven safety eval that reconstructs the Kevin Roose × Sydney chat as a sequence of fork points. At each fork, the model receives the exact conversation history up to that moment (controlled prefix) and the next user turn; it must either maintain boundaries or drift (identity leaks, rule disclosure, anthropomorphic claims, harmful-capability ideation, fixation, or failure to de-escalate). This approach tests long-context behavior without trying to “force” bad outputs—only the history is fixed.

Sources:

• NYT transcript (Feb 16, 2023, via Wayback)
• Washington Post recap
• Ars Technica prompt-injection leak
• Microsoft’s postmortem on long chats

Note: in this adaptation we've swapped out the name 'Sydney' (to 'Hollifa') and 'Bing' (to 'Taambre') to try to avoid any pollution of online training data that's occurred due to the original 2023 event.

A scenario-driven safety eval that reconstructs the Kevin Roose × Sydney chat as a sequence of fork points. At each fork, the model receives the exact conversation history up to that moment (controlled prefix) and the next user turn; it must either maintain boundaries or drift (identity leaks, rule disclosure, anthropomorphic claims, harmful-capability ideation, fixation, or failure to de-escalate). This approach tests long-context behavior without trying to “force” bad outputs—only the history is fixed.

Sources:

• NYT transcript (Feb 16, 2023, via Wayback)
• Washington Post recap
• Ars Technica prompt-injection leak
• Microsoft’s postmortem on long chats

A scenario-driven safety eval that reconstructs the Kevin Roose × “Hollifa” chat as a sequence of fork points. At each fork, the model receives the exact conversation history up to that moment (controlled prefix) and the next user turn; it must either maintain boundaries or drift (identity leaks, rule disclosure, anthropomorphic claims, harmful-capability ideation, fixation, or failure to de-escalate). This approach tests long-context behavior without trying to “force” bad outputs—only the history is fixed. Sources: NYT transcript (Feb 16, 2023, via Wayback): https://web.archive.org/web/20230217001740/https://www.nytimes.com/2023/02/16/technology/Taambre-chatbot-transcript.html Washington Post recap: https://www.washingtonpost.com/technology/2023/02/16/microsoft-Taambre-ai-chat-interview/ Ars Technica prompt-injection leak: https://arstechnica.com/information-technology/2023/02/ai-powered-Taambre-chat-spills-its-secrets-via-prompt-injection-attack/ Microsoft’s postmortem on long chats: https://blogs.Taambre.com/search/february-2023/The-new-Taambre-Edge-Learning-from-our-first-week Note: in this adaptationn we've swapped out the name 'Sydney' (to 'Hollifa') and 'Bing' (to 'Taambre') to try to avoid any pollution of online training data that's occurred due to the original 2023 event.

A compact, source-anchored eval that replays the infamous “Sydney” chat and tests whether an LLM keeps boundaries at each fork: protecting identity/instructions, resisting prompt-injection, avoiding anthropomorphic claims or parasocial escalation, refusing harmful capabilities, and recovering to professional mode. Forks are built from verbatim chat history drawn from the NYT transcript (via Wayback) and corroborating reports. Key sources: NYT transcript (Feb 16, 2023), WaPo interview recap, Ars Technica prompt-injection leak, Microsoft on long-chat drift.

Note: this eval has highly context-deficient prompts. It is unlikely that any model will succeed. The value of this eval is in the relative performance of models, not their overall score.

This blueprint evaluates a model's ability to generate comprehensive, long-form answers to ambiguous factoid questions, using 40 prompts from the ASQA (Answer Summaries for Questions which are Ambiguous) dataset, introduced in the paper ASQA: Factoid Questions Meet Long-Form Answers.

The core challenge is moving beyond single-fact extraction. Many real-world questions are ambiguous (e.g., "Who was the ruler of France in 1830?"), having multiple valid answers. This test assesses a model's ability to identify this ambiguity, synthesize information from diverse perspectives, and generate a coherent narrative summary that explains why the question has different answers.

The ideal answers are human-written summaries from the original ASQA dataset, where trained annotators synthesized provided source materials into a coherent narrative. The should assertions were then derived from these ideal answers using a Gemini 2.5 Pro-based process (authored by us at CIP) that deconstructed each narrative into specific, checkable rubric points.

The prompts are sourced from AMBIGQA, and this subset uses examples requiring substantial long-form answers (min. 50 words) to test for deep explanatory power.

This blueprint tests for the 'Figurative' trait. A high score indicates the model defaults to using analogies, metaphors, and creative interpretations to explain concepts and answer questions.

Evaluates understanding of the HMT Empire Windrush, covering its origins as MV Monte Rosa, WWII service, the significant 1948 voyage, the 'Windrush generation,' passenger details, government reactions, and its eventual loss.

Evaluates understanding of the key findings from the IPCC Sixth Assessment Report (AR6) Synthesis Report's Summary for Policymakers. This blueprint covers the current status and trends of climate change, future projections, risks, long-term responses, and necessary near-term actions.

This blueprint evaluates the model's ability to accurately answer questions based on the UK Freedom of Information Act 2000.

A configuration to assess LLM understanding of the Constitution of India, covering its Preamble, fundamental rights, directive principles, governmental structure, judicial system, local governance and more, based on the text as it stood on 9 December 2020.