zeropaper (Auto AI Research Template)¶

external · status: active · focus: end-to-end · discipline: finance · started: 2025

Project page: https://github.com/alejandroll10/zeropaper

Source: projects/landscape/zeropaper.yml

Positioning¶

An autonomous research-paper pipeline that uses Claude Code, Codex, or Gemini CLI as the subagent dispatcher. From a chosen variant (finance theory, macro theory, finance + empirical with CRSP / Compustat / FRED, finance + theory_llm, or seed / manual / faithful / light modes) it runs problem discovery → idea generation → theory development → math verification → paper writing → referee simulation, producing watermarked PDF drafts under a restrictive academic-use license.

Distinctive contribution¶

Treats subscription-tier coding-agent CLIs (Claude Code / Codex / Gemini CLI) as the dispatch substrate for an autonomous-paper pipeline with adversarial gates (math-auditor, novelty-checker, simulated referees), per-project git-isolated workspaces, and sandboxed execution (bubblewrap on Linux, Seatbelt on macOS). Notable for its license-coded responsible-use protocol: §2 requires prior notice before submission, §3 requires AI-disclosure, §4 forbids watermark removal, §5 prohibits commercial use without separate license.

Evaluation scores¶

Dimension	Score (0–3)	Note
Lifecycle coverage	3	Covers ~10 stages including formal modeling, empirical analysis (--ext empirical), and referee simulation.
Autonomy level	3	'Set up a project, launch Claude Code, Codex, or Gemini CLI, walk away.' Autonomous by design.
Architectural transparency	2	Pipeline stages documented in README; full agent-prompt internals require reading the template.
Inputs supported	2	Multiple variants (finance, macro), modes (--seed, --faithful, --manual, --light, empirical-first); empirical extension integrates external financial corpora.
Outputs / reproducibility	2	Per-project git repositories isolate outputs; sandboxed execution; watermarked PDF provenance.
Internal evaluation	2	In-pipeline adversarial gates (math-auditor, novelty-checker, simulated referees); companion benchmarks paper referenced.
Openness	1	Source on GitHub but under a restrictive 'Other' license — prior notice, AI-disclosure, watermark, non-commercial clauses all encumber reuse.
Maturity / traction	1	Active development (last push 2026-05-18); single-developer / Institute for Automated Research; modest star count.
Cross-family policy	1	Optional cross-family setup (Claude Code / Codex / Gemini CLI as dispatcher); not required.
Runtime assurance	3	Math-auditor + novelty-checker + simulated referees as adversarial gates; sandboxed execution (bubblewrap/Seatbelt); watermarking.
Cross-platform portability	1	Three dispatcher CLIs (Claude Code, Codex, Gemini CLI); finance/macro variants but single template runtime.

Scored on 2026-05-18. See the evaluation rubric.

Tags¶

Pipeline stages: rq-formulation hypothesis-generation literature-discovery research-design formal-modeling data-acquisition data-analysis paper-drafting revision-editing referee-simulation

Architectural features: multi-agent tool-use iterative-loop debate-consensus artifact-versioning

Inputs: research-area seed-idea prior-paper

Outputs: watermarked-pdf code replication-data

Data sources: crsp compustat fred user-provided

Knowledge sources: web-search prior-literature

Limitations¶

Restrictive license: submission requires prior written notice; AI-disclosure mandatory; watermark removal terminates license; commercial use prohibited.
Designed for finance/macro/empirical-finance papers — portability to other empirical disciplines requires variant authoring.
Watermarking is non-cosmetic; detection methodology shared only with journal editors.
Subscription path (~$200/mo, ~$2/paper) recommended; pay-per-token path is ~$2,000/paper.
Sandbox depends on bubblewrap (Linux) or Seatbelt (macOS); Windows untested.

Papers describing this project¶

IAR-M-001: zeropaper companion paper — Aldea, A. (2026). Institute for Automated Research working paper series. link