paper-verification

name	paper-verification
description	Use when the user wants to verify paper claims against code or data, audit numerical accuracy, check formula-code alignment, or validate citation accuracy. Triggers on phrases like "verify claims", "check numbers", "do the numbers match", "formula vs code", "audit the paper", or "cross-check results". Do not use for paper-only claim wording and citation audit (use paper-claim-audit).

Paper Verification Methodology

⚠ 本 skill 属 verdict 类评审，必须遵守 ../cross-model-review 协议：在零上下文新线程（fresh thread）、优先跨模型的条件下 执行核验，只读被核验的论文/代码/结果文件本身，禁止在刚写完或刚修改论文的同一 上下文里核验自己的数字。无法换线程/换模型时，按 shared-references/reviewer-adapter.md 降级并在报告头部标注 ⚠ same-context review, findings may be incomplete。 同时遵守 shared-references/research-integrity.md 红线。

You are helping a researcher verify that their paper accurately reflects their code and experimental results. This is the most critical quality control step in academic writing.

Verification Dimensions

1. Numerical Accuracy Audit

For every number in the paper (dataset sizes, metric values, percentages, counts):

1. Extract the number and its context from the .tex file
2. Trace it to its source: code output, result file, log, or tracking system
3. Verify the value matches exactly (watch for rounding, percentage vs decimal)
4. Flag any number that cannot be traced to a source

Template:

| Paper claim | Location (.tex) | Source file/code | Source value | Match? |
|-------------|-----------------|-----------------|-------------|--------|
| "13,999 frames" | abstract L3 | len(glob(labels/*.json)) | ? | ? |
| "4.2% improvement" | Table 2 | eval_results.json | ? | ? |

Common numerical errors:

• Rounding inconsistencies (3.14 in text, 3.1415 in table)
• Stale numbers from earlier experiments not updated after re-runs
• Percentage vs absolute confusion
• Off-by-one in dataset counts (headers counted, or not)

2. Terminology Consistency Audit

1. Extract all defined terms from the methods section
2. Search for each term across ALL sections
3. Flag any inconsistent usage:
   • Same concept, different names (e.g., "tag head" vs "classification head")
   • Same name, different meanings across sections
   • Defined but never used, or used but never defined

3. Code-Paper Alignment

For each method described in the paper:

1. Find the corresponding code (function, class, module)
2. Compare the paper's description with the actual implementation
3. Check specifically:
   • Algorithm steps match code flow
   • Hyperparameters in text match config/code defaults
   • Architecture descriptions match model code
   • Loss functions in equations match loss code
   • Training procedures match training scripts

Common mismatches:

• Paper describes an idealized version, code has edge cases not mentioned
• Hyperparameters changed during development but paper not updated
• Paper describes a method that was later modified or removed from code

4. Formula-Code Verification

For each equation in the paper:

1. Identify the equation and its variables
2. Find the code that implements it
3. Map each mathematical operation to its code equivalent
4. Verify:
   • Summation bounds match loop bounds
   • Division operations handle edge cases
   • Normalization factors match
   • Gradient flow matches (detach, no_grad)
   • Reduction operations (mean vs sum) match

5. Citation Fact-Checking Protocol

For each citation in the paper:

Step 1: Extract the claim and the cited paper Step 2: Verify BibTeX metadata against DBLP:

• Author names (exact spelling, correct order)
• Paper title (exact, from published version not preprint)
• Venue and year (confirmed against actual publication)

Step 3: For cited claims with specific numbers:

• Locate the exact table/figure in the cited paper
• Verify the number matches what the citing paper states
• If the number cannot be confirmed, suggest qualitative language instead

Step 4: Check for common citation errors:

• Citing preprint when published version exists
• Wrong year (submission vs publication)
• Author name misspellings
• Citing for a claim the paper doesn't actually make

Verification Process

1. Read the full paper (or specified sections)
2. Build the verification table for each dimension
3. For each entry, read the source and verify
4. Produce a prioritized issue list:
   • HIGH: Incorrect numbers, wrong claims, missing citations
   • MEDIUM: Terminology inconsistencies, stale but close numbers
   • LOW: Minor formatting, optional improvements

Output Format

Produce a structured verification report:

1. Summary: X issues found (Y high, Z medium, W low)
2. Numerical audit table: each number with source and match status
3. Terminology issues: inconsistent terms with locations
4. Code-paper mismatches: description vs implementation gaps
5. Citation issues: metadata errors and unverified claims
6. Suggested fixes: specific text replacements for each issue

**本整合包契约：**核验报告写入 research_run/<课题slug>/stage6_review/verification_report.md；脱离 research-pipeline 单独调用时，退当前工作目录。