Philosophical biases make AI code reviewers 75% more unique and zero false positives

A new arXiv paper from Kaushal Bansal introduces a radical approach to AI-assisted code review: instead of asking models to be generic 'expert reviewers,' the system constrains them with specific philosophical dispositions rooted in distinct epistemological traditions. The four dispositions tested are Pyrrhonist Skepticism (refuses to accept any claim without evidence), Navya-Nyaya logic (rigorous Indian logical analysis), Diogenes' Cynicism (aggressively questions utility and convention), and Confucian relational ethics (evaluates code for its impact on team harmony and maintainability). Each disposition is defined apophatically—by what it refuses to do—and equipped with a self-monitoring failure mode (called hamartia) to prevent blind spots. The dispositions are orchestrated in sequence by role protocols, producing structurally different types of findings.

Evaluated on 50 merged pull requests across 7 repositories spanning Python, Go, C++, Java, and Terraform, the system achieved 46% convergence with human reviewers (validating signal quality) while identifying unique findings at a 75% rate. Remarkably, across 601 total findings, not a single one was judged false-positive by the original author. A controlled baseline comparison showed that 51% of disposition findings were not produced by the same model using generic 'expert reviewer' prompting—and these unique findings targeted structural, operational, and logical concerns rather than standard code-level issues. Preliminary cross-model validation between Claude Opus and GPT Codex 5.3-xhigh on 3 PRs demonstrated 100% framework-structure adherence with only 39% finding-level agreement, suggesting the framework provides real behavioral constraint while preserving model-specific analytical perspective. The study spans pre-AI 2020 and post-AI 2024–2026 eras, covering both enterprise and open-source organizations.