1. Input and transcript
We analyze submitted text or a transcript obtained for a YouTube video. Transcript quality affects all later stages. Visuals, charts, tone, and on-screen disclosures are not independently reviewed.
2. Content classification
The existing classifier distinguishes informational, opinion, and entertainment content. Entertainment content may be marked not analyzed. Opinion can contain factual claims that are selected for review.
3. Claim selection
Up to six factual claims are selected and may be normalized to make them understandable outside the surrounding transcript. Selection is not exhaustive.
4. Retrieval planning
The worker builds searches for each selected claim using configured search and source connectors. Financial-authority coverage supplements general retrieval; it does not limit all reviews to financial topics.
5. Evidence retrieval
Connectors retrieve candidate sources. Coverage depends on indexing, access, service availability, and how precisely the claim is expressed.
6. Evidence assessment
Candidates are evaluated for relevance, authority, freshness, and overall quality. A retrieved source is not automatically usable evidence or support for a claim.
7. Claim findings and confidence
A finding describes the comparison: supported, likely true, mixed, unsupported, or inconclusive. Confidence is the model's confidence in that individual assessment; it is not a calibrated probability that the claim is true. When no usable evidence is available, the interface shows confidence as not assessed.
8. Coverage and summary
Coverage counts selected claims with usable evidence. It is separate from claim confidence and says nothing about unselected statements. Legacy numeric scores are heuristic summaries, not compliance ratings or guarantees.
9. Human review and limits
Check original sources, date-sensitive facts, context, and omissions. Missing evidence does not prove a claim false. AI can make mistakes. Accepted submissions count toward the plan allowance, including failures or entertainment classifications. Findings do not constitute legal advice or regulatory approval.