Chatgpt Factcheck

Workflow

Step 1: Crawl ChatGPT Conversation

Use Playwright to render the JavaScript-heavy ChatGPT shared page and extract content.

from playwright.sync_api import sync_playwright

with sync_playwright() as p:
    browser = p.chromium.launch(headless=True)
    page = browser.new_page()
    page.goto(URL, wait_until='networkidle', timeout=30000)
    page.wait_for_timeout(8000)

    # Extract conversation text
    full_text = page.inner_text('body')

    # Extract all source URLs (citations are in small badge-like <a> elements)
    links = page.eval_on_selector_all('a[href]',
        'els => els.map(e => ({text: e.innerText.trim(), href: e.href, class: e.className}))')

    # Filter for academic/data sources (exclude chatgpt.com, openai.com UI links)
    source_links = [l for l in links
        if l['href'] and 'chatgpt.com' not in l['href'] and 'openai.com' not in l['href']
        and any(domain in l['href'] for domain in [
            'zenodo', 'pubmed', 'sciencedirect', 'doi.org', 'arxiv',
            'springer', 'wiley', 'mdpi', 'nature', 'ieee', 'scopus',
            'github', 'nasa', 'nrel', 'nih.gov'
        ])]

    browser.close()

Fallback: If Playwright fails, ask user to paste conversation text directly.

Step 2: Extract Claims

From the conversation text, identify all verifiable factual claims:

• Dataset descriptions (size, format, license, content)
• Paper/publication references (title, authors, journal, year)
• Numerical claims (frequencies, pressures, measurements)
• Institutional attributions (who released what)
• URLs and DOIs

Create a structured claim list:

Claim #1: [claim text]
Source cited: [URL or "none"]
Category: dataset / publication / numerical / institutional

Step 3: Fact-Check Each Source

Visit each extracted URL to verify against the original source.

Strategy by domain:

Verdict categories:

• ✅ Confirmed — Source directly supports the claim
• ⚠️ Partial — Source exists but claim is exaggerated or slightly inaccurate
• ❌ Not found — URL is dead, or source does not support the claim
• 🔍 Unverifiable — No source cited, cannot independently confirm

Step 4: Analyze Paper Relevance

Read the target paper note from the vault to understand its structure and research gaps.

Default target: 03-projects/(Research) Paper-1 Review Multiscale Fatigue AI DT.md

Relevance mapping process:

1. Read the target paper note (frontmatter, chapter structure, key themes)
2. For each verified finding, check:
   • Which chapter/section does it relate to?
   • Does it fill a gap (missing reference, dataset, validation)?
   • Usage type: benchmark data / literature reference / method comparison / validation target
3. Also check Paper-2, Paper-3, Paper-4 relevance if connection is strong

Step 5: Generate Obsidian Note

Save to 01-inbox/ following vault conventions. Write note body in Korean with English technical terms in parentheses.

Filename: (Resource) [Descriptive Title].md

Frontmatter template:

---