What problem does it solve? Mathematical and statistical code often contains subtle bugs—missing normalization factors, off-by-one indexing, matrix transposition errors, or numerical instability—that pass ordinary code review. This Skill systematically verifies implementations against formal LaTeX derivations and catches discrepancies before they corrupt results. ## Core Features & Use Cases - Formal Proof Drafting: Generates LaTeX derivations, variable-to-symbol mapping tables, and line-by-line code alignment for the target implementation. - Adversarial Red Teaming: Spawns an isolated subagent that writes empirical verification scripts (sympy/numpy with pure-Python fallbacks) and audits against a 5-point discrepancy checklist covering scale, indexing, matrix dimensions, numerical stability, and boundary assumptions. - Socratic Human Signoff: Conducts an interactive interview to confirm the user understands the mathematical trade-offs before approval. - Persistent Proof Notes: Exports a self-contained proof reference note with embedded test logs to an Obsidian Vault. - Use Case: After implementing a custom loss function or softmax layer, run /showproof to verify the code matches the math, catch a missing 1/N normalization factor, and archive the verified proof for future reference. ## Quick Start Ask the agent to run /showproof on the file containing your mathematical implementation to verify it against formal derivations.
