SKILL.md

--- frontmatter

name: code-audit
description: >
  Verify consistency between LaTeX paper and Python implementation.
  Checks hyperparameter matches, architecture alignment, loss function
  correctness, and algorithm step ordering. Use after implementation
  or when checking if code matches the paper.
allowed-tools: Read, Grep, Glob, Bash

Code Audit — LaTeX ↔ Python Consistency

Quick Scan Script

Run the automated check first:

bash

# 1. Hyperparameter match via CONFIG-SYNC markers
echo "=== HYPERPARAMETER CHECK ==="
grep "CONFIG-SYNC" workspace/paper/main.tex | while read line; do
    param=$(echo "$line" | grep -oP '\w+ = [\w.e-]+')
    param_name=$(echo "$param" | cut -d= -f1 | xargs)
    param_val=$(echo "$param" | cut -d= -f2 | xargs)
    code_val=$(grep "$param_name" workspace/src/config.py | grep -oP '= .+' | head -1)
    echo "  LaTeX: $param_name = $param_val | Code: $param_name $code_val"
done

# 2. Architecture dimension check
echo "=== ARCHITECTURE CHECK ==="
grep -n "hidden_dim\|input_dim\|output_dim\|num_layers\|num_heads" workspace/src/config.py
grep -n "d_h\|d_{in}\|d_{out}\|L =" workspace/paper/main.tex

Manual Verification Checklist

1. Hyperparameters (MUST all match)

Parameter	LaTeX Location	config.py Variable
Learning rate	Section IV	`learning_rate`
Batch size	Section IV	`batch_size`
Epochs	Section IV	`num_epochs`
Weight decay	Section IV	`weight_decay`
Architecture dims	Section III	various

2. Loss Function

Read the loss equation in main.tex and compare term-by-term with loss.py:

• Main loss term matches
• Regularization terms match
• Weighting coefficients match
• Reduction method (mean vs sum) matches

3. Algorithm Steps

Read Algorithm 1 in main.tex and compare with train.py:

• Step ordering matches
• Gradient computation matches
• Special procedures (warmup, scheduling, clipping) match

4. Model Architecture

Read Section III and compare with model.py:

• Layer count matches
• Activation functions match
• Skip connections / residual structures match
• Each custom layer has equation reference comment

Output Format

Write findings to workspace/logs/code_audit.json:

json

{
  "timestamp": "...",
  "iteration": N,
  "status": "PASS | WARN | CRITICAL",
  "mismatches": [
    {
      "parameter": "learning_rate",
      "latex_value": "1e-3",
      "code_value": "1e-4",
      "severity": "CRITICAL"
    }
  ],
  "architecture_match": true,
  "loss_match": true,
  "algorithm_match": true,
  "score": X
}

Any CRITICAL mismatch = immediate pipeline halt.