Skip to content

ZeroPaper: Enhanced AI-Assisted Research Template with Agentic Capabilities

Key Ideas

  • ZeroPaper is now released as an open template: clone the repo once, run setup.sh to scaffold a project, then say "Run the pipeline" inside Claude Code, OpenAI Codex CLI, or Gemini CLI. The pipeline runs autonomously through problem discovery → idea generation → theory → math verification → paper writing → simulated referee review.
  • A single Claude Code Max-tier subscription (~$200/month) supports roughly 100 papers/month, putting the amortized cost at ~$2 per paper. Pay-per-token API access works but is ~1000× more expensive at current rates (~$2,000/paper) because of heavy subagent dispatch.
  • Variants target different journals: --variant finance (JF, JFE, RFS) and --variant macro (AER, Econometrica, QJE, JPE, ReStud, JME). Extensions are additive: --ext empirical adds CRSP/Compustat/FRED/Ken French/Chen-Zimmerman/WRDS data agents; --ext theory_llm adds LLM-experiment agents via UF NaviGator.
  • Pipeline has 10 numbered stages and 6 adversarial gates: every substantive advance is gated by a fresh-context agent that has not seen the generator's reasoning, with at least two evaluators at math gates and three simulated referees at the publication gate.
  • Roughly 30 specialized agents: literature-scout, idea-generator, idea-reviewer, idea-prototyper, theory-generator, math-auditor, math-auditor-freeform, novelty-checker, theory-explorer, self-attacker, scorer, paper-writer, referee, style, six polish-* agents (consistency/formula/numerics/institutions/equilibria/bibliography), bib-verifier, scribe, and (with extensions) empiricist, empirics-auditor, experiment-designer, experiment-reviewer.
  • Additional modes: --manual ships the agents as a standalone research toolkit without the autonomous loop, --seed lets you drop existing notes/drafts into output/seed/ and have the pipeline triage entry, --light swaps Sonnet in for all subagents to cut cost.
  • License is restrictive even for open release: submission of any pipeline-produced work to a journal/preprint/conference requires prior written notice (60-day fallback), AI-disclosure is mandatory, watermarks are non-cosmetic and removal terminates the license, and commercial use is prohibited without a separate license. Share-alike applies to derivatives.
  • Safety defaults sandbox bash to the project folder, block SSH/AWS credential reads, and (on Linux) enforce restrictions via bubblewrap; on macOS Seatbelt is used.

Summary

This release moves ZeroPaper from a private collaboration-by-invitation system (see summaries/automated-research-finance) to a publicly clonable template. The same pipeline that produced the 22-paper gallery (see summaries/zeropaper-gallery) is now scaffoldable in a single shell command, and crucially is runtime-portable: the same pipeline state is consumed identically by Claude Code, OpenAI Codex CLI, and Gemini CLI through three parallel orchestration docs (CLAUDE.md, AGENTS.md, GEMINI.md). A session can be interrupted in one runtime and resumed in another.

The architectural commitments visible in the README are the practical surface of the design principles developed in the companion paper: state lives in a compact JSON file that every stage reads and writes (not in conversation history); the orchestrator only routes — actual work is dispatched to fresh-context subagents; every gate runs adversarial verifiers with distinct framings; and termination is mechanical, not at the orchestrator's discretion. The dashboard at localhost:8000/dashboard.html exposes stage transitions and gate verdicts live, and the pipeline commits to git at every transition, giving researchers an auditable trail.

For an empirical-finance user, the most consequential piece is the --ext empirical extension, which bundles skills for SEC EDGAR, FRED, Ken French, Chen-Zimmerman's Open Source Asset Pricing signals, mutual-fund holdings, flexible empirical specification mining, and full WRDS access (CRSP, Compustat, IBES, options, insider trading). This is the same data perimeter Lopez-Lira used to generate the gallery papers, now exposed for arbitrary new projects.

Relevance to Economics Research

ZeroPaper is the first widely accessible autonomous-research pipeline targeting top finance and macro journals where the entire toolchain is documented, runnable on consumer-tier AI subscriptions, and tested against published-paper benchmarks. Three points matter for economists evaluating these systems:

  1. Cost structure is no longer the binding constraint. At ~$2/paper amortized on a $200/month subscription, the marginal cost of an additional pipeline run is effectively zero. This shifts the bottleneck from compute budget to idea quality, research-question novelty, and human review capacity — exactly the substitution the summaries/thread-arindube post predicts.
  2. The license terms preview the publishing-side fight. Mandatory submission notice, AI-disclosure, and non-removable watermarks operationalize the disclosure regime that journals will need if they want to track AI-produced submissions. Researchers using ZeroPaper inherit these obligations whether or not their target journal has a policy yet.
  3. The runtime-portability claim is testable. Because the same state JSON drives all three CLIs, ZeroPaper provides a controlled setting for measuring how much of "AI research quality" depends on model choice vs. orchestration. This connects to concepts/multi-agent-systems and to the design-principle arguments in the companion paper.

For users of summaries/ai-powered-scholarship, summaries/dickerson-ai-asset-pricing, summaries/ralph-wiggum-asset-pricing, and summaries/project-ape, this is the most direct head-to-head reference implementation now available for finance research pipelines.