Scores Alone Do Not Prove Discovery: The Discovery Certification Protocol for Auditing AI Research Agents

TL;DR

DCP protocol audits AI research agents, ensuring result recoverability and feedback effects, with zero recovery rate upper bound at 0.0468.

cs.MA 🔴 Advanced 2026-09-07 4 views
Jingjie Ning Shanshan Zhong Xiaochuan Li Ji Zeng
AI research protocol verification experimental design feedback effects recoverability

Key Findings

Methodology

The DCP protocol audits AI research agents through three gates. Gate 1 validates useful improvement on sealed evaluation; Gate 2 tests matched agents' recovery ability without historical data; Gate 3 measures the average effect of truthful feedback relative to a neutral policy. Experiments were conducted in SQLite optimization and virtual catalyst control, ensuring result recoverability and feedback effects.

Key Results

  • Zero recoveries in 96 episodes, with an upper bound of 0.0468.
  • 30 truthful recoveries and zero neutral recoveries, with passing 60-pair null studies.
  • Experiments verified the DCP protocol's effectiveness across different models.

Significance

The DCP protocol provides a standardized evidence language for AI research, ensuring result recoverability and feedback effects. This is significant for academia and industry, addressing long-standing verification challenges.

Technical Contribution

The DCP protocol introduces independent verification of recovery witnesses and feedback effects, providing technical assurance distinct from existing methods. It ensures research result reproducibility and verifiability, promoting transparency in AI research.

Novelty

The DCP protocol is the first to separate discovery verification from feedback effects, offering a common evidence language, unlike previous score-only verification methods.

Limitations

  • The protocol's complexity may increase implementation difficulty.
  • A large amount of experimental data is needed to verify the protocol's effectiveness.

Future Work

Future work could extend the DCP protocol to more AI research fields, exploring its applicability in different scenarios.

AI Executive Summary

The DCP protocol provides a systematic approach to auditing AI research agents' discoveries. Through three gates, DCP ensures the recoverability and feedback effects of research results. Gate 1 validates useful improvement on sealed evaluation, Gate 2 tests matched agents' recovery ability without historical data, and Gate 3 measures the average effect of truthful feedback relative to a neutral policy. Experiments were conducted in SQLite optimization and virtual catalyst control, showing zero recoveries in 96 episodes, with an upper bound of 0.0468. The DCP protocol provides a standardized evidence language for AI research, addressing long-standing verification challenges and is significant for academia and industry. While the protocol's complexity may increase implementation difficulty, it offers crucial assurance for transparency and verifiability in AI research.

Deep Analysis

Background

AI research agents have made significant advances in algorithm discovery and research automation. However, verifying these agents' discoveries has been challenging. Traditional methods often rely on scores, but these do not prove the authenticity and reproducibility of discoveries.

Core Problem

The core problem is how to verify AI research agents' discoveries. Scores alone cannot prove authenticity, lacking a systematic method to verify result recoverability and feedback effects.

Innovation

The DCP protocol systematically audits AI research agents' discoveries through three gates. Gate 1 validates useful improvement on sealed evaluation; Gate 2 tests matched agents' recovery ability without historical data; Gate 3 measures the average effect of truthful feedback relative to a neutral policy.

Methodology

  • �� Gate 1: Validate useful improvement on sealed evaluation. • Gate 2: Test matched agents' recovery ability without historical data. • Gate 3: Measure the average effect of truthful feedback relative to a neutral policy.

Experiments

Experiments were conducted in SQLite optimization and virtual catalyst control. Each experiment set clear targets and control conditions, ensuring result recoverability and feedback effects.

Results

Experiments showed zero recoveries in 96 episodes, with an upper bound of 0.0468. There were 30 truthful recoveries and zero neutral recoveries, with passing 60-pair null studies.

Applications

The DCP protocol can be used to audit AI research agents' discoveries, particularly in scenarios requiring result recoverability and feedback effects.

Limitations & Outlook

The protocol's complexity may increase implementation difficulty. Additionally, a large amount of experimental data is needed to verify the protocol's effectiveness.

Plain Language Accessible to non-experts

Imagine you're cooking in a kitchen. The DCP protocol is like a strict recipe that not only tells you what ingredients you need but also requires you to document how you do each step. This way, even if you're not in the kitchen, others can make the same dish using this recipe. This protocol ensures the repeatability of each step and the verifiability of the results.

ELI14 Explained like you're 14

Imagine you're playing a game where you need to complete a task to level up. But this game has a special rule: you not only have to complete the task but also prove how you did it. The DCP protocol is like this rule in the game, ensuring every player can prove their task completion method, so everyone competes fairly.

Glossary

DCP Protocol

A systematic method for auditing AI research agents' discoveries.

Used to verify result recoverability and feedback effects.

Recovery Witness

Evidence proving the recoverability of research results.

Used in Gate 2 to verify matched agents' recovery ability.

Feedback Effect

The average effect of truthful feedback relative to a neutral policy.

Evaluated in Gate 3 to assess feedback effectiveness.

Sealed Evaluation

Testing conducted without disclosing historical data.

Used to verify the authenticity of research results.

Matched Agent

An agent tested for recovery ability without historical data.

Used to verify the recoverability of research results.

Open Questions Unanswered questions from this research

  • 1 How can the DCP protocol be applied to broader AI research fields?
  • 2 How does the DCP protocol perform in real-time applications?

Applications

Immediate Applications

AI Research Verification

The DCP protocol can be used to audit AI research agents' discoveries, ensuring result recoverability and feedback effects.

Long-term Vision

Research Transparency

Through the DCP protocol, transparency and verifiability in AI research will be significantly enhanced.

Abstract

AI research agents combine prior knowledge, public sources, and experimental feedback to produce useful results. The Discovery Certification Protocol (DCP) turns claims about these results into executable recovery and feedback tests. Gate 1 validates useful improvement on sealed evaluation. Gate 2 gives matched agents the registered starting information and observed Web content while withholding the target research history. Every valid method reaching the numerical target supplies a recovery witness and triggers the Core veto. DCP Core requires adequate controls, zero observed recoveries, and a finite-sample bound on recovery in one fresh registered episode. Optional Gate 3 measures the average effect of truthful feedback relative to a specified neutral policy from a shared checkpoint. DCP Evidence adds this effect after independent null calibration and a registered effect margin. Two controlled audits exercise the complete protocol in SQLite optimization and virtual catalyst control under different models. Each produced zero recoveries in 96 episodes, with an upper bound of 0.0468. Each paired study yielded 30 truthful recoveries and zero neutral recoveries, with passing 60-pair null studies. Additional cases exercise Core, recovered, and audit-incomplete decisions. A deterministic, LLM-free verifier reproduces the decisions from frozen evidence. DCP provides a common evidence language for useful outcomes, alternative routes, and feedback effects across AI research.

cs.MA cs.AI cs.SE