Problem framing
Translate a phenomenon, mechanism intuition, or target claim into a research object with clear scope.
Projects
Projects extend the lab's research agenda beyond individual papers.
Each project outlines its substantive question, current research stage, related themes, and participating collaborators.
AI-native workflow
Adrasteia supports economic paper writing through structured assumption search.
The workflow is based on a preliminary, under-review manuscript about human-in-the-loop agentic research for theory-oriented economics. It is presented here as a project overview rather than as a public manuscript release.
Its purpose is not to automate paper writing end to end. Instead, it helps researchers turn early ideas into inspectable models, branch across alternative assumptions, revise arguments under internal checks, and prepare drafts for referee-style scrutiny.
How it helps economic paper writing
- Sharper research questions
- More defensible assumptions
- Traceable assumption-to-conclusion chains
- Drafts that better survive reviewer-style criticism
Assumption proposal
Generate candidate assumption packages and model branches that can support a disciplined economic argument.
Internal verification
Check logic, derivations, literature fit, boundary conditions, and whether conclusions outrun assumptions.
Revision and branching
Weaken claims, substitute mechanisms, refine assumptions, and preserve promising alternatives for comparison.
Human adjudication
Let the researcher select, prune, or retain branches based on originality, taste, and scholarly accountability.
Simulated reviewer validation
Expose draft research objects to referee-style critique, then convert feedback into structured revision targets.
Active
Adrasteia: AI Agent Workflow for Economic Paper Writing
Accepted at the ICML 2026 Workshop on AI for Math, the EC'26 Workshop on AI-Driven Research in EconCS (oral contributed talk), and the STOC 2026 TheoryFest workshop "Can AI Do Theory?" A human-in-the-loop agentic workflow for moving economic research from an initial question to defensible assumptions, model structure, argument revision, and referee-style critique.
Team — Heikichi Hayashi
Design
The Elite School Custodianship Effect
A work-in-progress project on the impact of super high school group-based schooling on students' human capital, developed in collaboration with Beijing National Day School.
Team — Heikichi Hayashi
Data
Primary data holdings support the lab's empirical work.
Corpora the lab has assembled and maintains for its own research, spanning judicial records, original fieldwork, and hand-built statistical panels.
Text corpus
Chinese Court Judgment Research Corpus, 2014–2024
A geocoded corpus built from the monthly public-release files of China's national judgment archive: 4.9 million criminal judgments in seventeen offense categories and 8.0 million civil documents in seven cause-of-action families, mapped to 330 prefectures and aggregated into monthly panels. Release files for 2021 to 2024 were retained before bulk public access to the archive closed and are no longer collectible from public sources. Dictionary-coded text measures are validated against a 4,624-document audited gold standard with independent blind human recoding.
- Coverage
- 2014–2020 monthly panels, plus retained 2021–2024 release files
- Scale
- 12.8 million judgments across 330 prefectures
- Provenance
- Bulk public releases preserved before archive access closed; in-house geocoding, text coding, and measurement audits
Qualitative fieldwork
Field Interview Corpus on Informally Governed Markets
Original fieldwork with participants in informally governed markets: 99 individual and 10 group interview sessions conducted over 2021–2023 and 2025–2026 with market operators, agents, players, police officers, a procurator, and a lender, concentrated in Guangdong, Jiangxi, and Fujian with telephone interviews elsewhere. Transcripts are double-coded into six mechanism themes, yielding 573 deduplicated verbatim extracts across 69 sessions. Respondents are identified by role and codebook ID only.
- Coverage
- Interview waves in 2021–2023 and 2025–2026
- Scale
- 109 sessions; 573 coded verbatim extracts across six themes
- Provenance
- Lab-conducted fieldwork under a respondent-protection protocol
Hand-collected panel
City-Level Gig Platform Entry Panel
A hand-collected record of when app-based gig platforms entered each Chinese city: platform arrival dates and the earliest archived rider-recruitment and crowdsourcing postings for 369 prefecture-level and provincial cities over the 2013–2018 rollout of food-delivery work. The panel dates each city's exposure to gig employment and supports staggered-adoption designs on household labor allocation.
- Coverage
- 369 cities over the 2013–2018 platform rollout
- Scale
- 369 cities with multiple dated entry markers each
- Provenance
- Hand-collected from platform records, corporate announcements, and recruitment archives
Hand-digitized panel
Hand-Digitized Small-Loan Company Provincial Panel
A province-by-quarter panel of licensed small-loan company counts, staffing, paid-in capital, and outstanding loan balances, digitized from the People's Bank of China regional statistical tables for 2014 to 2020. Most quarters are parsed from PDF attachments; periods whose tables survive only as image scans were coded by hand. The consolidated panel is not available in machine-readable form from any public source.
- Coverage
- 31 provincial units, 2014Q3 to 2020Q4
- Scale
- 744 province-quarter observations over 24 quarters
- Provenance
- Parsed and hand-coded from scattered PBoC statistical report attachments