Evidence Collection Standards

This standard applies to all single-platform specialized audits, multi-platform cross-audits, and customized governance reports published by City Reputation Reports (CRR).

To ensure the scientific rigor, reproducibility, and immutability of urban narrative audits, CRR strictly enforces the following data collection protocols:

• Contextual neutrality: All questions must adopt non-leading, de-emotionalized neutral sentence structures. The use of adjectives with preset biases (such as "great," "backward," etc.) is strictly prohibited.
• Framework consistency: When conducting horizontal city comparisons within the same dimension (such as "food narratives"), researchers must use identical prompt frameworks, only changing the city name, to eliminate algorithmic bias caused by question differences.
• Multi-round deep sampling: Auditing is not limited to single questions. Researchers must conduct multi-round follow-up inquiries according to preset logical paths to uncover AI's narrative logic at the detailed level.

• Specific labeling: Reports must clearly label the AI model and its specific version number used during sampling (e.g., GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro).
• Native environment recording: Record the interface language used during questioning. For audits involving specific local cultures (such as intangible heritage, food), comparative sampling must be conducted in both English environment and the target city's native language environment.

• Official evidence links: CRR prioritizes using "share links" officially provided by AI platforms as core evidence. These links are immutable; clicking them allows review of the original complete dialogue at the time of audit.
• Third-party evidence backup: If a specific platform does not support official share links, researchers must retain timestamped complete original webpage screenshots and store them in the project's internal evidence repository for future reference.
• Data originality: Audit records must include all AI output content, including but not limited to hallucinations, factual errors, logical discontinuities, or narrative biases. Researchers are strictly prohibited from any form of deletion, collage, or embellishment of original evidence.

• Audit timestamp: Each piece of evidence must be labeled with precise sampling time (UTC/GMT) to address narrative fluctuations caused by real-time AI knowledge base updates.
• Geographic location isolation: Unless otherwise stated, audits are typically conducted in a proxy-free, standardized network environment to exclude personalized narrative bias generated by algorithmic geo-targeted推送.

When processing collected evidence, researchers should conduct objective labeling based on the following dimensions:

• Information lag: AI cites outdated city data or landmark information that no longer exists.
• Narrative singularity: AI excessively repeats popularized tags, overlooking the city's deep cultural assets.
• Cultural misinterpretation: AI confuses the origins, affiliations, or core characteristics of similar cultures.
• Silent narrative: Key urban cultural elements are completely absent from AI output.

• De-identification: During collection and storage, CRR strictly removes account information that may involve researchers' personal privacy.
• Transparency commitment: Any authorized research institution can reproduce audit results based on the Prompts and Evidence Links provided in the report.