AI evidence extraction

ZeroPaper: An Autonomous Research System

Alejandro Lopez-Lira · Open MIND · 2026

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10

Relevance

1/4

Quality (LMQS)

Evidence

Citations

0.00

FWCI

This is an AI analysis. Read the peer-reviewed original at the publisher: https://doi.org/10.5281/zenodo.20127842

Methodology & findings

Study design

Architectural design and systems engineering.

Primary method

Design Science (architecture-patterns approach inspired by Alexander et al. 1977, Gamma et al. 1994, Hohpe and Woolf 2003). The design derives from ten premises about LLM behavior and deployment economics, formalized into six design principles that serve as a pattern language for auditing system architecture.

Main result

The pipeline is structured to avoid the failure modes most often degrading automated-research prototypes. The system executes ten numbered stages and six adversarial gates, coordinating roughly thirty specialized agents through a versioned artifact tree. "The narrower claim, the one this paper develops, is that the pipeline is structured to avoid the failure modes most often degrading automated-research prototypes, and that the principles it implements are the ones those failure modes call for." The best produced papers would receive a revise-and-resubmit at a finance field journal under blinded review, with output volume of roughly one hundred finished papers per month per subscription at the average run length.

Research paradigm

Design Science / Systems Engineering

Author conclusions

"ZeroPaper is an autonomous pipeline for end-to-end research paper production. The system executes ten numbered stages and six adversarial gates, coordinating roughly thirty specialized agents through a versioned artifact tree, under any of three host runtimes, at around two dollars per paper amortized under a flat-fee max subscription." The author concludes that the three central design choices (adversarial isolation of evaluators, mechanical termination preventing infinite loops, and multi-evaluator gates with distinct framings) combined with multi-runtime packaging keep the pipeline portable. "Not every run produces a publishable paper. The pipeline's structure avoids the failure modes I have most often seen degrade automated-research prototypes, but whether the surviving papers clear a real journal's bar is a question the system cannot answer from inside itself. Submission is the only test."

Risk of bias

Self-bias in LLM-generated content (Premise 1); Potential mechanism drift between claimed and actual mathematical mechanisms; Model-level training biases inherited across agents (Premise 8); Potential specification gaming in agent outputs (Premise 5); Long-context degradation affecting model judgment (Premise 2); Coherence drift across multiple stages (Premise 3); Path-of-least-resistance bias where models prefer shortcuts (Premise 5); Single evaluator blind spot matching generator's blind spot; Mechanism drift where claimed mechanism may not match mathematical delivery; Self-bias in LLMs defending prior output (Premise 1); Generator-evaluator shared blind spots despite isolation; Model-level correlation in fresh instances (Premise 8); Specification gaming and path-of-least-resistance behavior (Premise 5); Temperature-based non-determinism across runs

Limitations

"Output quality: calibrated but not externally validated
While the finance variant of the pipeline is designed to target the top-three finance journals (JF, JFE, RFS), my conjecture based on finance-refereeing experience is that the best of the finance-variant papers it has produced would receive a revise-and-resubmit at a peer-reviewed finance field journal such as the Journal of Financial and Quantitative Analysis under blinded review." Additional limitations include: non-deterministic output across runs making baseline comparison difficult, paywalled full-text citations reducing semantic-fit verification, mechanism drift detection limitations with single adversarial agent, and "Dependency on host runtime behavior" where "the pipeline's behavior inherits from the host runtime's behavior (Claude Code, Codex, or Gemini)
When a host updates its subagent semantics, tool availability, or system-prompt handling, the runtime-specific packaging layer has to track the change."

Open questions raised

Mechanism drift detection remains an open problem
Paywalled full-text semantic-fit checking is not yet implemented; currently bounded to abstract-level verification
External validation of output quality by independent referees
Comparison protocols for non-deterministic autonomous systems against non-deterministic baselines
Automated tracking of host runtime behavior changes
External validation through actual journal submission and peer review

Data: The empirical extension adds access to financial databases (CRSP, Compustat, FRED, WRDS) through a long-running local socket server.; Academic graph data via OpenAlex (for bibliography verification); Financial databases accessible through extension: CRSP, Compustat, FRED, WRDS; Financial databases accessible through empirical extension: CRSP, Compustat, FRED, WRDSCode: https://github.com/alejandroll10/zeropaper (main system repository); https://github.com/alejandroll10/zeropaper-gallery (companion gallery of autonomously generated papers); https://github.com/alejandroll10/zeropaper-gallery (gallery of autonomously generated papers); https://github.com/alejandroll10/zeropaper; https://github.com/alejandroll10/zeropaper-galleryExtracted from: pdfAgreement 55%

Explore related topics

Systematic Review Automation Research Productivity Design Science