At Referee, we not only expect the system to be gamed, we encourage it! Why? It stems from our belief that adversarial processes make the strongest systems. By encouraging and anticipating how the system can be gamed early, we hope to build a robust system that engenders trust across all stakeholders. After all, we take a hacker’s mindset to the problem of research evaluation, so why not embrace the approach to our own system?
This is a living document that will continue to add possible ways the system can be gamed and how we would approach them.
Collusive Reviewing
- Technique: A group of validators might collude to consistently approve each other’s bounty claims, regardless of their validity.
- Solution: Implement a random assignment of validators to claims and introduce periodic audits of validator activities by independent auditors.
Sybil Attacks
- Technique: An individual could create multiple accounts to influence the system, either to push through fraudulent claims or to manipulate reputation scores.
- Solution: Require robust identity verification processes for each account to ensure one person cannot control multiple accounts.
Pseudo-Flaw Flooding
- Technique: Flooding the system with trivial or borderline irrelevant flaws to claim bounties, thus diluting the focus on more significant issues.
- Solution: Develop a tiered system for flaws based on their impact on the paper’s integrity, adjusting bounty rewards accordingly.
Reputation Laundering
- Technique: Validators or bounty hunters artificially inflate their reputation through minor but numerous successful bounty claims.
- Solution: Introduce weighting mechanisms that give more importance to the impact of each claim on scientific discourse rather than just the number of successful claims.
Echo Chamber Creation
- Technique: A partisan group could dominate certain topics or fields, pushing their biased views by selectively approving bounties that align with their stance.
- Solution: Ensure diversity in the validator pool and set up mechanisms that require consensus among validators from different backgrounds.
Bounty Claim Hoarding
- Technique: Bounty hunters could hoard information about significant flaws to claim multiple bounties before others notice.
- Solution: Set a time limit for holding off on public disclosure after the initial discovery of a flaw, encouraging timely reporting.
Bribery and Corruption
- Technique: Bounty providers could bribe validators to approve claims on their papers, or validators could solicit bribes in exchange for claim approval.
- Solution: Implement strict anti-corruption policies and penalties, including permanent bans for bribery.
Data Poisoning
- Technique: Introducing subtly incorrect data or references in papers that can mislead automated scanners and validators.
- Solution: Combine automated tools with expert human review to catch nuanced manipulations not detected by AI.
Targeted Smear Campaigns
- Technique: Using the platform to systematically lower the reliability scores of researchers, institutions or ideologies through coordinated attacks.
- Solution: Monitor for abnormal patterns in bounty submissions and validations, possibly integrating a peer-review model among validators to check each other’s work.
Exploitation of Model Weaknesses
- Technique: Identifying and exploiting specific weaknesses in the AI models used for initial paper scans to manipulate review outcomes.
- Solution: Continuously update and test AI models against known exploitation techniques and encourage the community to report new vulnerabilities.