Audit-Closed AI Scientist: A Benchmark for Statistically Valid Autonomous Scientific Discovery

The idea of an AI scientist is no longer speculative. Recent systems show that AI agents can generate research ideas, run experiments, produce figures, and even write scientific papers. Systems such as The AI Scientist demonstrate an end-to-end research loop: idea generation, experimentation, visualization, and automated paper writing. (sakana.ai)

At the same time, this emerging paradigm raises a fundamental question:

How do we ensure that discoveries produced by autonomous research systems are statistically valid?

A new open-source benchmark attempts to address this problem:

GitHub repository: https://github.com/kadubon/audit-closed-ai-scientist

The project introduces an experimental framework called the Audit-Closed AI Scientist protocol, designed to test whether autonomous research agents maintain statistical integrity under adaptive experimentation.

The emerging era of AI scientists

Automation of science has been a long-standing goal in AI. Early projects such as Dendral and later Robot Scientist systems aimed to automate hypothesis generation and experimental testing.

Today, large language models and autonomous agent architectures are pushing this idea much further. Modern AI scientist systems can:

• generate research hypotheses
• search literature
• write experimental code
• run experiments
• generate plots
• produce full research manuscripts

Several recent systems have demonstrated these capabilities, including autonomous research frameworks that can generate peer-reviewed scientific papers or propose new hypotheses across multiple domains. (arXiv)

These systems effectively transform the scientific process into a closed loop of automated discovery.

However, there is a critical challenge.

The statistical problem of autonomous discovery

When research is automated and scaled, statistical failure modes become easier to trigger.

Autonomous systems can rapidly explore thousands of hypotheses, continuously monitor results, and stop experiments as soon as results appear significant. This introduces classical statistical pitfalls such as:

• p-hacking
• optional stopping
• hypothesis shopping
• adaptive experiment selection

These problems are well known in statistics and sequential inference research. Methods such as e-values and confidence sequences were developed precisely to maintain statistical validity under continuous monitoring and adaptive experimentation. (stat.cmu.edu)

In a large-scale autonomous research pipeline, these issues become even more important.

Without safeguards, an AI scientist could easily generate convincing but unreliable discoveries simply by exploring enough hypotheses.

The Audit-Closed AI Scientist protocol

The repository introduces a simple but strict principle:

Accept_t = f(Log_0:t)

In other words:

Scientific acceptance decisions must depend only on a publicly verifiable log of experiments.

Every hypothesis, experiment, and evaluation step is recorded in a transparency log. The log functions as an auditable record of the research process.

Evidence accumulation is implemented using sequential e-process inference, which remains statistically valid even when:

• experiments are monitored continuously
• hypotheses are generated adaptively
• experiment selection depends on previous results

This approach turns the research pipeline into an audit-closed process: decisions are determined entirely by the experiment history.

What the repository provides

The repository is not an AI scientist system itself.

Instead, it provides a benchmark environment for evaluating autonomous research pipelines.

The benchmark includes:

1. Baseline simulations

Naive AI scientist pipelines that generate hypotheses and run experiments adaptively.

These simulations reproduce classical failure modes such as:

• p-hacking
• optional stopping
• candidate shopping

2. Audit-closed protocol implementation

An alternative research pipeline using:

• transparency logs
• deterministic replay
• sequential e-process inference

3. Comparative benchmarks

Experiments measuring:

• false discovery rate
• replication probability
• statistical stability under adaptive experimentation

The benchmark allows researchers to directly compare naive autonomous discovery pipelines with audit-closed protocols.

Why this matters

Autonomous research agents are becoming increasingly capable.

Systems are already being used for:

• automated machine learning research
• materials discovery
• biomedical hypothesis generation
• AI-assisted literature synthesis

As these systems scale, the reliability of automated discoveries becomes critical.

A research pipeline that can explore thousands of hypotheses per day must also ensure that its conclusions remain statistically valid.

The Audit-Closed AI Scientist benchmark focuses on precisely this problem.

It provides a testbed for evaluating statistical integrity in autonomous research systems.

Intended audience

This benchmark may be useful for researchers working on:

• AI scientist systems
• autonomous research agents
• automated scientific discovery
• self-driving laboratories
• scientific machine learning infrastructure

It may also interest researchers studying:

• reproducible science
• statistical validity in adaptive experimentation
• AI governance for research automation

Reproducing the experiments

All simulations in the repository are reproducible.

git clone https://github.com/kadubon/audit-closed-ai-scientist
pip install -r requirements.txt
python run_all_experiments.py

The repository includes scripts to regenerate benchmark results and figures.

Benchmark Results: Statistical Validity in AI Scientist Systems

To evaluate the statistical reliability of AI scientist pipelines and autonomous research agents, the benchmark simulates common adaptive-experimentation failure modes and compares them with the Audit-Closed protocol.

Key Empirical Findings

1. Many-Hypothesis Inflation

Under the global null (no true signal):

| Hypotheses tested | Naive pipeline FDR | Bonferroni | | — -| — -| — -| | 5 | 0.193 | 0.0267 | | 1000 | 1.000 | 0.050 |

Observation: large hypothesis search alone can produce false discoveries in naive automated research pipelines.

2. Candidate Design Shopping

False-positive rate under repeated experiment design selection:

| Candidate designs | Naive pipeline | Sequential e-process | | — -| — -| — -| | 5 | 0.134 | 0.000 | | 10 | 0.237 | 0.000 | | 25 | 0.511 | 0.000 | | 50 | 0.717 | 0.000 |

Observation: adaptive experiment design selection strongly inflates false discoveries in naive pipelines.

3. Optional Stopping

Continuous monitoring with 400 sequential looks:

| Method | False-positive rate | | — -| — -| | Peeking p-values | 0.339 | | Fixed-horizon p-values | 0.0425 | | Sequential e-values | 0.0367 |

Observation: sequential e-process inference remains close to nominal calibration under adaptive monitoring.

4. Integrated Autonomous Research Benchmark

Comparison of two simulated pipelines:

| Metric | Baseline AI scientist | Audit-Closed protocol | | — -| — -| — -| | False discovery rate | 0.653 | 0.000 (95% CI upper bound 0.0119) | | Replication success | 0.722 | 0.809 |

Observation: the audit-closed protocol substantially reduces false discoveries while maintaining comparable replication reliability in this benchmark environment.

5. Adversarial Experiment Submission

Simulated attack with 100 malicious candidate experiments:

| Metric | Result | | — -| — -| | Baseline false acceptance | 1.000 | | Audit-Closed false acceptance | 0.002 | | Replay verification match | 1.000 | | Tamper detection | 1.000 |

Observation: transparency logs and deterministic replay significantly improve robustness in this adversarial simulation.

Scope

This benchmark evaluates statistical validity in autonomous scientific discovery pipelines, including optional stopping, multiplicity control, and adversarial experiment submission. It should be interpreted as a simulation framework for evaluating AI scientist research systems, not a full autonomous scientist implementation.

Toward trustworthy autonomous science

Autonomous scientific discovery is advancing rapidly.

From early expert systems to modern AI scientists, the trajectory suggests that machines will play an increasingly central role in generating scientific knowledge.

But with that capability comes a new requirement:

autonomous science must also be auditable science.

Benchmarks such as the Audit-Closed AI Scientist repository are small steps toward that goal.

They help us ask a simple but essential question:

When an AI claims to have discovered something, how do we know the discovery is statistically trustworthy?