AgentSkillsCN

check-tables

核对 LaTeX 表格与文中引用内容之间的数值一致性。对于包含大量结果表格的论文而言，这一点至关重要。

SKILL.md
--- frontmatter
name: check-tables
description: Verify numerical consistency between LaTeX tables and the prose that references them. Critical for papers with many result tables.
allowed-tools: Read, Grep, Glob, Bash

Table-Text Consistency Checker

Systematically verify that every number in the paper's prose matches its source table.

Step 1: Extract Tables

For each \begin{table} ... \end{table} environment:

  1. •Record the table label (from \label{})
  2. •Record the caption text
  3. •Extract all numerical values and their column/row context
  4. •Note the table's position in the document

Step 2: Find Prose References

For each table label, find every \ref{tab:...} and read the surrounding paragraph. Also search for numbers that appear in both a table and nearby prose without explicit \ref.

Step 3: Cross-Check

For each number mentioned in prose that relates to a table:

CRITICAL mismatches:

  • •Number in text differs from number in table (e.g., text says "23.0%" but table has "22.8%")
  • •Claimed improvement doesn't match actual difference (e.g., "improves by 3.2%" but 85.1 - 82.4 = 2.7)
  • •Superlatives that are wrong ("best performance" but another row is higher)
  • •Ranking claims that are wrong ("second-best" but it's actually third)

WARNING:

  • •Number appears in text but no table reference nearby
  • •Table referenced but no specific numbers discussed (might be fine, might be lazy)
  • •Rounding inconsistency (table has 82.37, text says 82.4 in one place and 82.37 in another)
  • •Different decimal precision for same metric across tables

INFO:

  • •Tables that are never referenced in text (orphaned tables)
  • •Very large tables where automated checking may miss context

Step 4: Validate Computed Values

If tables contain values that are computed from other values:

  • •Averages: verify they match the individual values
  • •Deltas/improvements: verify subtraction is correct
  • •Percentages: verify they sum appropriately (to 100% where expected)
  • •Bold/underline: verify "best" markers match actual best values

Output Format

code
TABLE CONSISTENCY REPORT
========================

Table 1 (tab:main_results) — "Main Results"
  ✅ Line 142: "82.3% accuracy" matches Table 1, row 3, col "Acc"
  ❌ Line 148: "improves by 3.2% over baseline" — actual difference is 2.7%
     Table value: 82.3, Baseline: 79.6, Difference: 2.7 (not 3.2)
  ⚠️  Line 155: "best performance on SQuAD" — Table 1 shows row 5 is higher

Table 2 (tab:ablation) — "Ablation Study"
  ✅ All 4 referenced numbers match
  ⚠️  Table 2 is never referenced in the text

Summary: N tables, N numbers checked, N critical, N warnings