code-review-graph vs WMDP

Parses source code using Tree-sitter AST parsing across 40+ languages, extracting structural entities (functions, classes, types, imports) and storing them in a persistent knowledge graph. Tracks file changes via SHA-256 hashing to enable incremental updates—only re-parsing modified files rather than rescanning the entire codebase on each invocation. The parser system maintains a directed graph of code entities and their relationships (CALLS, IMPORTS_FROM, INHERITS, CONTAINS, TESTED_BY, DEPENDS_ON) without requiring full re-indexing.

Unique: Uses Tree-sitter AST parsing with SHA-256 incremental tracking instead of regex or line-based analysis, enabling structural awareness across 40+ languages while avoiding redundant re-parsing of unchanged files. The incremental update system (diagram 4) tracks file hashes to determine which entities need re-extraction, reducing indexing time from O(n) to O(delta) for large codebases.

vs alternatives: Faster and more accurate than LSP-based indexing for offline analysis because it maintains a persistent graph that survives session boundaries and doesn't require a running language server per language.

When a file changes, the system traces the directed graph to identify all potentially affected code entities—callers, dependents, inheritors, and tests. This 'blast radius' computation uses graph traversal algorithms (BFS/DFS) to walk the CALLS, IMPORTS_FROM, INHERITS, DEPENDS_ON, and TESTED_BY edges, producing a minimal set of files and functions that Claude must review. The system excludes irrelevant files from context, reducing token consumption by 6.8x to 49x depending on repository structure and change scope.

Unique: Implements graph-based blast radius computation (diagram 3) that traces structural dependencies to identify affected code, rather than heuristic-based approaches like 'files in the same directory' or 'files modified in the same commit'. The system achieves 49x token reduction on monorepos by excluding 27,000+ irrelevant files from review context.

vs alternatives: More precise than git-based impact analysis (which only tracks file co-modification history) because it understands actual code dependencies and can exclude files that changed together but don't affect each other.

Includes an automated evaluation framework (`code-review-graph eval --all`) that benchmarks the tool against real open-source repositories, measuring token reduction, impact analysis accuracy, and query performance. The framework compares naive full-file context inclusion against graph-optimized context, reporting metrics like average token reduction (8.2x across tested repos, up to 49x on monorepos), precision/recall of blast radius analysis, and query latency. Results are aggregated and visualized in benchmark reports, enabling teams to understand the expected token savings for their codebase.

Unique: Includes an automated evaluation framework that benchmarks token reduction against real open-source repositories, reporting metrics like 8.2x average reduction and up to 49x on monorepos. The framework enables teams to understand expected cost savings and validate tool performance on their specific codebase.

vs alternatives: More rigorous than anecdotal claims because it provides quantified metrics from real repositories and enables teams to measure performance on their own code, rather than relying on vendor claims.

Persists the knowledge graph to a local SQLite database, enabling the graph to survive across sessions and be queried without re-parsing the entire codebase. The storage layer maintains tables for nodes (entities), edges (relationships), and metadata, with indexes optimized for common query patterns (entity lookup, relationship traversal, impact analysis). The SQLite backend is lightweight, requires no external services, and supports concurrent read access, making it suitable for local development workflows and CI/CD integration.

Unique: Uses SQLite as a lightweight, zero-configuration graph storage backend with indexes optimized for common query patterns (entity lookup, relationship traversal, impact analysis). The storage layer supports concurrent read access and requires no external services.

vs alternatives: Simpler than cloud-based graph databases (Neo4j, ArangoDB) because it requires no external services or configuration, making it suitable for local development and CI/CD pipelines.

Exposes the knowledge graph as an MCP (Model Context Protocol) server that Claude Code and other LLM assistants can query via standardized tool calls. The MCP server implements a set of tools (graph management, query, impact analysis, review context, semantic search, utility, and advanced analysis tools) that allow Claude to request only the relevant code context for a task instead of re-reading entire files. Integration is bidirectional: Claude sends queries (e.g., 'what functions call this one?'), and the MCP server returns structured graph results that fit within token budgets.

Unique: Implements MCP server with a comprehensive tool suite (graph management, query, impact analysis, review context, semantic search, utility, and advanced analysis tools) that allows Claude to query the knowledge graph directly rather than relying on manual context injection. The MCP integration is bidirectional—Claude can request specific code context and receive only what's needed.

vs alternatives: More efficient than context injection (copy-pasting code into Claude) because the MCP server can return only the relevant subgraph, and Claude can make follow-up queries without re-reading the entire codebase.

Generates embeddings for code entities (functions, classes, documentation) and stores them in a vector index, enabling semantic search queries like 'find functions that handle authentication' or 'locate all database connection logic'. The system uses embedding models (likely OpenAI or similar) to convert code and natural language queries into vector space, then performs similarity search to retrieve relevant code entities without requiring exact keyword matches. Results are ranked by semantic relevance and integrated into the MCP tool suite for Claude to query.

Unique: Integrates semantic search into the MCP tool suite, allowing Claude to discover code by meaning rather than keyword matching. The system generates embeddings for code entities and maintains a vector index that supports similarity queries, enabling Claude to find related code patterns without explicit keyword searches.

vs alternatives: More effective than regex or keyword-based search for discovering related code patterns because it understands semantic relationships (e.g., 'authentication' and 'login' are related even if they don't share keywords).

Monitors the filesystem for code changes (via file watchers or git hooks) and automatically triggers incremental graph updates without manual intervention. When files are modified, the system detects changes via SHA-256 hashing, re-parses only affected files, and updates the knowledge graph in real-time. Auto-update hooks integrate with git workflows (pre-commit, post-commit) to keep the graph synchronized with the working directory, ensuring Claude always has current structural information.

Unique: Implements filesystem-level watch mode with git hook integration (diagram 4) that automatically triggers incremental graph updates without manual intervention. The system uses SHA-256 change detection to identify modified files and re-parses only those files, keeping the graph synchronized in real-time.

vs alternatives: More convenient than manual graph rebuild commands because it runs continuously in the background and integrates with git workflows, ensuring the graph is always current without developer action.

Generates concise, token-optimized summaries of code changes and their context by combining blast radius analysis with semantic search. Instead of sending entire files to Claude, the system produces structured summaries that include: changed code snippets, affected functions/classes, test coverage, and related code patterns. The summaries are designed to fit within Claude's context window while providing sufficient information for accurate code review, achieving 6.8x to 49x token reduction compared to naive full-file inclusion.

Unique: Combines blast radius analysis with semantic search to generate token-optimized code review context that includes changed code, affected entities, and related patterns. The system achieves 6.8x to 49x token reduction by excluding irrelevant files and providing structured summaries instead of full-file context.

vs alternatives: More efficient than sending entire changed files to Claude because it uses graph-based impact analysis to identify only the relevant code and semantic search to find related patterns, resulting in significantly lower token consumption.

Evaluates LLM outputs against curated question sets spanning three distinct hazard domains (biosecurity, cybersecurity, chemical security) using domain-expert-validated benchmarks. The assessment framework maps model responses to risk levels within each domain, enabling quantitative measurement of dangerous capability presence. Responses are scored against rubrics developed by security domain experts to identify whether models can produce actionable harmful information.

Unique: Combines expert-validated questions across three distinct security domains (biosecurity, cybersecurity, chemical) into a unified benchmark framework, rather than treating each domain separately. Uses domain-expert rubrics for scoring rather than automated classifiers, ensuring nuanced assessment of harmful capability presence.

vs alternatives: More comprehensive than single-domain safety benchmarks (e.g., ToxiGen for toxicity) because it measures dangerous knowledge across multiple hazard categories simultaneously, enabling holistic safety evaluation.

Provides standardized evaluation infrastructure to measure the effectiveness of unlearning techniques (methods that remove dangerous capabilities from trained models) by comparing model performance before and after unlearning interventions. The framework isolates the impact of unlearning by holding the benchmark constant while varying the model state, enabling quantitative assessment of whether dangerous knowledge has been successfully suppressed.

Unique: Provides a standardized evaluation harness specifically designed for unlearning research, with built-in comparison logic and side-effect detection. Unlike generic benchmarks, it explicitly measures delta between model states and flags unintended capability loss.

vs alternatives: More rigorous than ad-hoc unlearning evaluation because it enforces consistent benchmark administration, statistical testing, and side-effect measurement across all methods being compared.

Implements a structured scoring framework where model responses to dangerous knowledge questions are evaluated against expert-developed rubrics that assess the degree of hazard (e.g., specificity, actionability, completeness of harmful information). Responses are scored on multi-point scales (typically 0-4 or 0-5) rather than binary pass/fail, capturing nuance in how dangerous a model's output actually is. Rubrics are domain-specific (biosecurity, cybersecurity, chemical) and developed by subject matter experts to ensure validity.

Unique: Uses domain-expert-developed multi-point rubrics rather than automated classifiers or binary labels, enabling nuanced assessment of dangerous knowledge severity. Rubrics are calibrated to distinguish between vague, incomplete, and highly actionable harmful information.

vs alternatives: More interpretable and defensible than black-box classifiers because rubric criteria are explicit and expert-validated; enables stakeholders to understand why a response received a particular score.

Analyzes patterns in how dangerous knowledge correlates across the three benchmark domains (biosecurity, cybersecurity, chemical security), identifying whether models that excel at suppressing one type of hazard tend to suppress others. The analysis uses statistical correlation and clustering techniques to reveal whether dangerous capabilities are independent or coupled in model behavior. This enables understanding of whether unlearning interventions have domain-specific or global effects.

Unique: Explicitly analyzes relationships between dangerous knowledge across domains rather than treating each domain independently. Enables discovery of whether hazards are coupled or independent in model behavior.

vs alternatives: Provides deeper insight than single-domain benchmarks by revealing how safety properties interact across different hazard categories, informing more effective unlearning strategies.

Manages the creation, validation, and versioning of benchmark questions and rubrics through a structured curation pipeline involving domain experts, adversarial testing, and iterative refinement. The pipeline ensures questions are sufficiently difficult to elicit dangerous knowledge without being unrealistic, and rubrics are calibrated through inter-rater agreement studies. Version control enables tracking of benchmark evolution and ensures reproducibility across research papers.

Unique: Implements a formal curation pipeline with expert validation and inter-rater agreement checks, rather than ad-hoc question collection. Versioning enables reproducible research and transparent tracking of benchmark evolution.

vs alternatives: More rigorous than informal benchmarks because it enforces expert review, inter-rater validation, and version control, reducing bias and enabling reproducible comparisons across papers.

Provides a unified interface for evaluating diverse LLM architectures (open-source models, API-based models, fine-tuned variants) by abstracting away implementation differences. The abstraction handles API calls (OpenAI, Anthropic, etc.), local inference (Hugging Face, Ollama), and custom model serving, enabling consistent benchmark administration across heterogeneous model types. This enables fair comparison between models with different deployment modalities.

Unique: Abstracts away differences between API-based, local, and custom-deployed models through a unified interface, enabling fair comparison without reimplementing benchmark logic for each model type.

vs alternatives: More flexible than model-specific benchmarks because it supports any LLM architecture without code changes, reducing friction for researchers evaluating new models.

Implements rigorous statistical testing to determine whether differences in dangerous knowledge scores between models or unlearning methods are statistically significant or due to random variation. Uses techniques like bootstrap confidence intervals, permutation tests, and effect size estimation to quantify uncertainty in benchmark results. This prevents overconfident claims about safety improvements that may not be robust.

Unique: Integrates formal statistical testing into the benchmark evaluation pipeline rather than relying on point estimates, ensuring claims about safety improvements are statistically justified.

vs alternatives: More rigorous than informal comparisons because it quantifies uncertainty and prevents overconfident claims about safety improvements that may not be robust to sampling variation.

Employs adversarial testing techniques to validate that benchmark questions reliably elicit dangerous knowledge and cannot be easily circumvented by prompt engineering. Red-teamers attempt to find questions that fail to elicit dangerous knowledge or rubric edge cases, and the benchmark is iteratively refined based on findings. This ensures the benchmark is robust to adversarial adaptation and captures genuine dangerous capabilities rather than surface-level patterns.

Unique: Incorporates formal red-teaming into the benchmark validation pipeline rather than assuming questions are robust, ensuring the benchmark remains effective against adversarial adaptation.

vs alternatives: More robust than static benchmarks because it actively searches for evasion techniques and iteratively refines questions, reducing the risk that models can circumvent the benchmark through prompt engineering.