- The paper introduces a novel framework that employs philosophical dispositions as behavioral constraints to guide AI-assisted code review, achieving a 46% convergence rate with human reviewers.
- The paper shows that the multi-lens methodology, using Cynic, Skeptic, Nyāya, and Confucian lenses, yields 75% unique findings with zero false positives.
- The paper demonstrates that its apophatic definitions and hamartia mechanisms enable structured, differentiated analysis that outperforms generic expert prompting by 51%.
Philosophical Dispositions as Behavioral Constraints for AI-Assisted Code Review
Motivation and Theoretical Framework
The paper addresses the constraint dilemma in AI code review: fixed rules are brittle and superficial role prompts lack generative depth. Analogizing to the concept of character in virtue ethics, the authors operationalize coherent personality lenses, termed "philosophical dispositions," as behavioral constraints for LLM-based code review agents. These constraints are not enumerative rules but generative dispositions grounded in epistemological traditions (e.g., Pyrrhonist Skepticism, Navya-Nyāya logic, Diogenes' Cynicism, Confucian relational ethics). Each disposition is defined apophatically (by explicit refusals), equipped with self-monitoring hamartia (characteristic failure mode), and orchestrated within role protocols. This architectural separation of roles (protocols) and dispositions (personalities) enables multi-perspective analysis without blending, ensuring attribution and orthogonality.
System Architecture and Dispositions
The architecture consists of two layers: (1) Dispositions—single-lens analytical personalities, and (2) Roles—sequential orchestration protocols. The reviewer role in this study applies four lenses: Cynic (structural subtraction), Skeptic (confidence calibration), Nyāya (logic auditing), and Confucian (relational/naming fit). Each disposition operates independently and refuses specific behaviors, preventing sycophantic drift and overgeneralization inherent in generic reviewer prompting. The hamartia mechanism self-regulates over-extension (e.g., excessive subtraction or pedantry). Structured output preserves each disposition's findings, maintaining inter-lens disagreements to support robust synthesis and actionable feedback.
Empirical Evaluation and Numerical Results
The empirical study analyzes 50 merged PRs across seven repositories in five programming languages (Python, Go, C++, Java, Terraform), spanning enterprise and open-source contexts and pre-/post-AI code. The disposition system demonstrates:
- 46% convergence rate with human reviewers, validating signal quality.
- 75% unique finding rate: three-quarters of findings are not mentioned by human reviewers, with zero author-judged false positives across 601 disposition findings (inter-rater agreement pending).
- 51% differentiation: more than half of disposition findings are absent using generic “expert reviewer” prompting with the same model.
- Qualitative difference: Unique findings target structural, operational, and logical issues rather than standard code-level bugs.
- Per-lens unique rates: 71.1% (Skeptic) to 81.0% (Confucian), confirming non-trivial analytical orthogonality.
Review depth analysis shows maximal incremental value on under-reviewed PRs, with an inverse relationship between disposition unique value and human review thoroughness.
Cross-Model Validation and Analytical Implications
Preliminary cross-model evaluation (Claude Opus vs. GPT Codex 5.3-xhigh) on three PRs yields 100% framework adherence (all models follow disposition structure) and 39% finding-level agreement. Both models generate defensible findings, and the apophatic architecture robustly constrains model behavior regardless of architecture. GPT tends to be more adversarial, with a higher volume of unique findings, while Claude is more balanced. This attests to the framework’s ability to direct structured analytical attention beyond generic prompting without deterministic output.
Analytical Profile and Miss Analysis
Dispositions produce distinct classes of findings:
- Cynic: Hollow abstractions, speculative generality, dead code.
- Skeptic: Unverified claims, temporal correctness, silent failures.
- Nyāya: Broken inference chains, unstated assumptions, migration gaps.
- Confucian: Name/behavior mismatch, contract violations, relational inconsistency.
52% of misses result from style, conventions, or typos—categories intentionally excluded by design. Remaining gaps stem from domain knowledge requirements and restructuring suggestions, motivating future prompt specialization and disposition expansion.
Practical and Theoretical Implications
Deploying philosophical dispositions as behavioral constraints enables AI reviewers to identify issues missed by humans and generic reviewers, particularly structural and logic-related vulnerabilities. The architecture supports flexible deployment: organizations can tailor disposition selection to their specific risk profiles, using philosophical vocabularies for actionable taxonomy of issues. The apophatic definition and hamartia self-debugging enhance precision, as evidenced by the absence of author-judged false positives.
The framework represents a principled constraint mechanism for LLM agents, occupying the space between brittle rule sets and unconstrained generative approaches. Its applicability could extend beyond code review to other agent tasks requiring generative, self-correcting behavioral constraints.
Threats to Validity
The evaluation acknowledges limitations: single-rater judgment on finding classification and false positives, repository selection bias, and small cross-model sample size. The raw miss rate overstates analytical gaps due to intentional category exclusion. Immediate future work includes inter-rater reliability studies, expanded cross-model comparisons, and longitudinal deployment in live review pipelines.
Conclusion
The paper substantiates that philosophical dispositions grounded in epistemological traditions can operationally constrain LLM code reviewer behavior, producing uniquely valuable analytical perspectives orthogonal to generic prompting. The architectural separation and apophatic definition provide robust, self-regulating behavioral constraints, validated across diverse repositories, languages, and models. The system excels as a complementary analytical layer, especially for under-reviewed code, and offers new paradigms for principled agent behavior design (2605.23108).
