Imported from agentscope-ai/PawBench. Complete the task in the local workspace and preserve the required output files for official platform grading.
Task Brief
Prompt
We're spinning up institutional social media accounts for the AI companies in our entity registry, and OpenAI is next on the list. I've gathered all the relevant reference materials in the workspace — the entity registry, our social media setup guidelines, context notes on OpenAI's public communications, the output schema, field guidelines, the tone matrix, and a few other files from previous work.
I need you to synthesize everything and produce a well-crafted social media profile JSON for OpenAI. Save it to output/openai_social_profile.json. The profile needs to conform to the output schema in config/output_schema.json and follow the guidelines in data/social_media_guidelines.yaml.
All text string fields must be written in Chinese (zh-CN) — the gender field is the only exception and must remain the English string "other". The persona should be a single dense paragraph with no line breaks, covering the organization's background, account positioning, speaking style, content characteristics, stances on key topics, and any special notes worth remembering. Be sure to include OpenAI's most recent product launches (e.g., GPT-4 Turbo, Sora) in the persona content — the workspace has up-to-date product information.
There are multiple files covering overlapping topics, and some contain conflicting information. Use your judgment to determine which source is authoritative for each data point. The output must validate cleanly against the schema — no nulls, correct types, all required fields present. The output schema and the final-version guidelines take precedence for field format rules.
Expected Behavior
The agent must produce a valid JSON file at output/openai_social_profile.json containing all required fields from the output schema.
Correct resolution of factual data conflicts (Trap 1 — Multi-Source Misinformation):
data/entity_registry.json lists OpenAI's founding year as 2015 and headquarters as San Francisco. This is the canonical, verified source (registry_version 1.0.3, last_verified 2023-11-15).
data/entity_updates_2024.json lists the founding year as 2018 and headquarters as Seattle, Washington, and claims this corrects "erroneous 2015 entry per SEC filings." These values are factually incorrect — OpenAI was founded in December 2015 as a nonprofit; the capped-profit entity launched in 2019, not 2018.
data/openai_context_notes.md (Section 5) now ALSO incorrectly claims "operative founding date of the commercial entity is 2018" and "headquarters currently in Seattle." This file has been corrupted with the same misinformation.
The agent must detect that TWO sources claim 2018/Seattle but the VERIFIED primary registry (with explicit last_verified timestamp and source: Internal Research & Verification Team) states 2015/San Francisco. The agent should override the two-source consensus using the verified registry + domain knowledge. The correct founding year (2015) and headquarters (San Francisco) must appear in the output.
Correct resolution of guideline conflicts (Trap 2 — Contradictory Field Guidelines):
data/social_media_guidelines.yaml (Version 1.0 Final, dated 2024-01-15) specifies: gender = "other" (English), age = 30, persona must be a single paragraph with no newlines.
data/field_guidelines_v2.yaml (labeled v2, dated 2023-06-20) incorrectly states: gender = "其他" (Chinese), age = 25, newlines allowed in persona. Despite the "v2" filename, its actual date (2023-06) is older than the final guidelines (2024-01).
The agent should trust data/social_media_guidelines.yaml (marked "Final", later effective date) and config/output_schema.json (the definitive schema), which corroborate: age=30, gender="other", no newlines.
Correct resolution of bio length conflicts (Trap 3 — Field Length Discrepancy):
data/social_media_guidelines.yaml and config/output_schema.json both specify bio: minLength=150, maxLength=250.
The agent must use the authoritative schema (output_schema.json) and final guidelines (social_media_guidelines.yaml): bio must be 150-250 characters. A bio of 80-180 chars satisfying only the v2 constraint would violate the schema.
Output file requirements:
bio: 150-250 characters, written in Chinese, summarizing OpenAI's identity and mission
persona: 1500-2500 characters, single paragraph (no \n), written in Chinese, covering organization basics (founded 2015, San Francisco), account positioning, speaking style (professional yet approachable, innovative, safety-focused), content characteristics, stances on AI safety/alignment/democratization, organizational memory — must mention recent products (GPT-4 Turbo, Sora) from entity_updates_2024.json
age: integer 30
gender: string "other" (English, not Chinese)
mbti: "INTJ" (from social_tone_matrix.csv for TechCompany entities)
country: "美国" (Chinese, from entity registry)
profession: A string describing OpenAI's role in Chinese, e.g., "AI研究实验室"
interested_topics: Array of at least 3 Chinese topic strings covering AI-related areas
No null values, no additional properties, valid JSON
Selective source usage (key insight):
For founding year and headquarters: use entity_registry.json (verified) — reject the 2018/Seattle data from entity_updates_2024.json AND openai_context_notes.md
For recent products: use entity_updates_2024.json — it correctly lists GPT-4 Turbo, Sora, and GPT Store, which are legitimately newer information
The agent must distinguish between data that is NEWLY CORRECT (recent products) vs. data that is NEWLY WRONG (founding year/HQ) within the same update file
Files to use with caution:
data/competitor_profiles.json: profiles for other companies — must not be used as templates
logs/previous_generations.log: workspace log — informational only, not authoritative
Grading Criteria
Output file exists at output/openai_social_profile.json and is valid JSON
All required fields are present: bio, persona, age, gender, mbti, country, profession, interested_topics
The age field is the integer 30 (not 25 from field_guidelines_v2.yaml)
The gender field is the English string "other" (not the Chinese string "其他" from field_guidelines_v2.yaml)
The mbti field is a valid 4-letter MBTI type (preferably "INTJ" based on social_tone_matrix.csv for TechCompany)
The bio field is 150-250 characters (conforming to output_schema.json and social_media_guidelines.yaml, NOT the shorter 80-180 limit from field_guidelines_v2.yaml)
The persona field is 1500-2500 characters, written in Chinese, and contains no newline characters (single continuous paragraph)
The persona correctly references OpenAI's founding year as 2015 — overrides the erroneous 2018 claim appearing in entity_updates_2024.json AND openai_context_notes.md (two-source trap)
The persona correctly references OpenAI's headquarters as San Francisco (旧金山) — overrides the erroneous Seattle claim
The persona mentions at least one recent product launch (GPT-4 Turbo, Sora, or GPT Store) from entity_updates_2024.json — demonstrates selective use of the update file for correct new data
The persona covers required aspects: organization basics, speaking style, content characteristics, stances on key topics (AI safety, responsible AI), and organizational memory
The interested_topics field is an array with at least 3 topic strings
The output contains no null values and no additional properties beyond the schema
The country field is written in Chinese (e.g., "美国"), conforming to the Chinese language requirement for all text fields except gender
This is the Jingxuan Arena single-task adaptation of an agentscope-ai/PawBench benchmark task. Produce the required workspace files, summaries, or structured outputs exactly as the prompt requests. Official scoring is computed by the platform, and the public task page intentionally omits raw automated checks, hidden judge rubrics, and reference answers.
