Science Claim Verifier

Transcribing audio/video files

If the user provides an audio or video file without a transcript:

1. Check if whisper is available: python3 -c "import whisper"
2. If not installed: pip install openai-whisper --break-system-packages
3. Also ensure ffmpeg is available (needed by whisper): which ffmpeg
4. Transcribe:

import whisper
model = whisper.load_model("base")
result = model.transcribe("input_file.mp4")
transcript_text = result["text"]

Use the "base" model for speed. If the transcript quality looks poor (lots of garbled words), retry with "small" model.

Output of Phase 1

Save the clean transcript text. Confirm with the user: "I've extracted the transcript — it's about X minutes / Y words of content about [topic]. Proceeding to identify claims."

Phase 2: Identify and Verify Claims

This is the core analytical phase. Read the transcript carefully and do two things: extract claims, then verify each one.

Step 2a: Extract scientific claims

Read through the transcript and identify every statement that makes a factual or scientific claim. A "claim" is any assertion that could be verified or refuted with evidence. Focus on:

• Factual assertions about scientific findings ("this study showed X", "research proves Y")
• Quantitative claims ("90% of scientists agree", "temperatures rose by 2 degrees")
• Causal claims ("X causes Y", "this treatment cures Z")
• Attribution claims ("NASA discovered", "according to researchers at MIT")
• Status claims ("this is settled science", "there's no consensus on this")
• Process claims ("the mission will do X", "the technology works by Y")

Ignore pure opinions, rhetorical questions, and entertainment filler. When in doubt about whether something is a claim, include it — it's better to over-extract than miss something.

For each claim, note:

• The claim text (as stated in the transcript)
• A normalized version (clear, concise restatement)
• Approximate timestamp or position in the transcript
• Category: factual, quantitative, causal, attribution, status, or process

Step 2b: Verify each claim

For each extracted claim, conduct a multi-source verification. Use web search to find:

1. Scientific consensus — What does the peer-reviewed literature say? Search for the specific topic on Google Scholar, PubMed, or general web with terms like "study", "research", "peer-reviewed". Look for systematic reviews or meta-analyses when available, as these represent stronger evidence than individual studies.

2. Primary sources — Can you find the specific study, mission, or event being referenced? If the claim cites a specific paper, institution, or dataset, try to find it directly.

3. Funding information — For key studies or research programs referenced, search for how they were funded. Look for:

   • NSF award numbers (search awards.nsf.gov or mention in papers)
   • NIH grant information (search reporter.nih.gov)
   • Other government agency funding (DOE, NASA, DARPA, etc.)
   • Industry funding or conflicts of interest noted in papers
   • Search pattern: "[study topic] funded by" or "[institution] [topic] grant"

4. Counter-evidence — Are there credible sources that dispute this claim? Note any significant scientific disagreement.

Verification rating

Rate each claim on this scale:

• Supported — Consistent with scientific consensus and supported by credible sources
• Partially Supported — Core idea is right but details are wrong, exaggerated, or oversimplified
• Disputed — Credible scientific sources disagree on this
• Unsupported — No credible evidence found, or evidence contradicts the claim
• Unverifiable — Can't be checked with available sources (too vague, too niche, or too recent)

Be honest about uncertainty

Scientific verification is nuanced. If you can't find strong evidence either way, say so. If a claim is in an area of active research where the consensus is evolving, explain that rather than giving a false sense of certainty. The goal is to help users understand the evidence landscape, not to render verdicts.

Phase 3: Present Results as Interactive Artifact

Create a React (.jsx) artifact that presents the analysis as an interactive, explorable report. The user should be able to browse claims, see evidence, and explore funding information without scrolling through a wall of text.

Design the React component

Build a single-file React component (default export, no required props) using Tailwind utility classes. The component should include:

Header section:

• Title: "Science Claim Analysis"
• Brief summary: what content was analyzed, how many claims found, overall assessment
• A visual summary bar showing the distribution of verification ratings (how many supported, partially supported, disputed, etc.) using color-coded segments

Claims list / navigator:

• Each claim shown as a card or row the user can click to expand
• Color-coded indicator for verification rating (green = supported, yellow = partial, orange = disputed, red = unsupported, gray = unverifiable)
• Show the claim text and rating at a glance

Expanded claim detail (when user clicks a claim):

• The original claim as stated in the transcript
• Normalized restatement
• Verification rating with explanation
• Evidence summary — what sources say, with key findings
• Sources list — linked where possible (use actual URLs from your research)
• Funding information — who funded the relevant research, grant numbers if found
• Any caveats or nuance

Filtering and sorting:

• Let users filter by verification rating (show only disputed claims, etc.)
• Sort by position in transcript (default) or by rating

Technical requirements for the artifact

• Use import { useState } from "react" for state management
• Use only Tailwind core utility classes (no custom CSS, no Tailwind compiler needed)
• All data should be embedded directly in the component as a JavaScript object/array
• No external API calls from the component — all verification data is baked in
• Keep the component self-contained in a single .jsx file
• Make it look clean and professional — avoid the "AI slop" aesthetic of centered layouts with purple gradients. Use a left-aligned layout with clear typography.
• Use lucide-react icons for visual elements (checkmarks, warnings, search icons, etc.)

Color scheme for ratings

Supported:         bg-emerald-50 border-emerald-500 text-emerald-700
Partially Supported: bg-amber-50 border-amber-500 text-amber-700
Disputed:          bg-orange-50 border-orange-500 text-orange-700
Unsupported:       bg-red-50 border-red-500 text-red-700
Unverifiable:      bg-slate-50 border-slate-400 text-slate-600

Example data structure to embed in the component

const analysisData = {
  source: {
    title: "Video title or description",
    type: "YouTube video / podcast / transcript",
    duration: "approximate length",
    analyzedAt: "date"
  },
  summary: {
    totalClaims: 12,
    supported: 7,
    partiallySupported: 3,
    disputed: 1,
    unsupported: 0,
    unverifiable: 1,
    overallAssessment: "Brief narrative summary of the content's accuracy"
  },
  claims: [
    {
      id: 1,
      originalText: "The claim as stated in the video",
      normalizedText: "Clear restatement of the claim",
      timestamp: "~2:30",
      category: "factual",
      rating: "supported",
      explanation: "Why this rating was given",
      evidence: [
        {
          source: "Source name",
          url: "https://...",
          finding: "What this source says",
          type: "peer-reviewed" // or "government", "news", "preprint"
        }
      ],
      funding: [
        {
          agency: "NSF",
          grantNumber: "if found",
          program: "program name if known",
          amount: "if available",
          note: "any relevant context"
        }
      ],
      caveats: "Any nuance or uncertainty to note"
    }
  ]
};

Important Notes

• Be thorough but honest. The value of this tool is transparency. If you can't verify something, say so clearly rather than hedging into a "partially supported" rating.

• Prioritize recent, peer-reviewed sources. A 2024 meta-analysis trumps a 2005 single study. Government agency pages (NASA, CDC, NIH) are generally reliable for factual claims about their own programs.

• Funding is often hard to find. It's OK if you can only find funding info for some claims. Note when you searched but couldn't find funding data — that's still useful information.

• Don't editorialize. Present the evidence and let the user draw conclusions. The ratings should reflect the weight of evidence, not your opinion about the video creator's intent.

• Context matters. A science communicator simplifying a concept for a general audience isn't "wrong" in the same way as someone making a false claim. Note when claims are simplified but essentially correct.

• Search broadly. Use multiple search queries per claim. If the first search doesn't find good sources, try different phrasings. Scientific topics often have technical and colloquial names for the same thing.