Claims & evidence / Structure a conversation

Structure a conversation

Use argument analysis to extract candidate claims from messages and retain who said what, where it appeared, and what it may require as evidence.

Preserve message identity#

const result = await superb.analyzeArgument({
  messages: [
    { id: 'a', text: 'This can help.' },
    { id: 'b', text: 'This will help.', replyTo: 'a' },
  ],
});

HTTP equivalent: POST /v1/arguments/analyze. Supply 1–20 messages with unique IDs and no more than 40,000 total characters. Each message is limited to 12,000 characters. maxClaims may be 1–40 and defaults to 24.

Inspect before presenting#

claims contains sentence-based candidates with exact UTF-16 spans. relations identifies supported wording relationships. semanticMismatches contains candidate sense differences. verificationNeeds lists claims that may need outside evidence.

A sentence classified as factual is not a verified fact. The analysis does not score a person’s intelligence or choose a winner.

Keep private analysis explicit#

Analysis is not retained by default. store: true requires an authenticated contributor and stores the result privately. Use it only when the person has chosen to save the analysis.

Design the next question#

Present a disputed term, a changed commitment, or a source request that someone can actually answer. Keep the original messages accessible. That is a better foundation for Resolve than an opaque overall score.

Next: Evaluate the evidence attached to one candidate claim.