research-algorithm-checker

What problem does it solve?

It helps researchers rigorously verify that target code implements the intended research algorithm correctly, catching bugs, logic mismatches, numerical mistakes, boundary issues, and efficiency problems before they waste more training or experiment cycles.

Core Features & Use Cases

• Step-by-step algorithm walkthrough: Traces the algorithm from entry to outputs, confirming each transformation against the expected design.
• Boundary, shape, dtype, and edge-case analysis: Checks for indexing errors, null/empty handling, tensor shape/dtype mismatches, and other failure points.
• Numerical and computation correctness checks: Validates mathematical implementations (loss/probabilities/thresholds/gradient flow) for correctness.
• Performance and efficiency review: Identifies wasted computation, memory inefficiency, and cache/parallelization opportunities.
• Design-to-implementation consistency validation: Compares code against papers/notes/pseudocode to pinpoint deviations.
• Issue tracking with deduplication: Uses ISSUES.md to track todo/doing/done/discard so discarded optimization ideas are not repeated.

Quick Start

Use the research-algorithm-checker skill only when you can provide the exact trigger phrase “科研启动-检查”, then tell it which repository files/functions you want reviewed and what reference (paper/pseudocode/design) to align with.