SigWise
View as Markdown Open in ClaudeOpen in ChatGPT

User trust scores ​

A yes/no fraud flag tells you who to stop. A trust score tells you how far to let everyone else go: who can list without review, whose payouts clear instantly, who needs to verify first. SigWise keeps a score for every user, updated as they act, on a scale whose levels you describe in words.

Define the score ​

A score signal places each user on an ordered list of levels. The answer is a number from 0 (the first level) to levels − 1 (the last), with fractions in between, plus a confidence.

json
PUT /v1/signals/trust_score
{
  "type": "score",
  "instructions": "How much can other users and the platform trust this user?",
  "criteria": [
    "high risk: deceptive, abusive or clearly fraudulent",
    "unproven: new or thin history, nothing wrong yet",
    "neutral: normal activity, minor issues",
    "trusted: long, consistent, positive history"
  ]
}

Every new account already has a three-level trust_score; replace its levels with ones that match the decisions you make.

Feed it evidence ​

Trust is built from behaviour over time, so send the events a human reviewer would weigh:

json
POST /v1/objects/user-42/events
{
  "object_type": "user",
  "events": [
    { "type": "event", "name": "order.completed", "metadata": { "amount": 120, "rating": 5 } },
    { "type": "event", "name": "report.received", "metadata": { "reason": "no-show", "reporter_trust": "trusted" } },
    { "type": "message", "content": "Sorry, I was sick, can we reschedule for tomorrow?" }
  ]
}

Each analysis sees the user's recent events plus a rolling summary of everything before, so one bad day does not erase a long good record, and a long record does not hide a sudden change.

Read and use it ​

json
GET /v1/objects/user-42
{
  "object_id": "user-42",
  "analysis": [
    { "key": "trust_score", "type": "score", "score": 2.4, "confidence": 0.8 }
  ]
}
  • Gate actions in your code. Instant payouts above 2.5, manual review below 1.
  • Rank users. GET /v1/objects?sort=trust lists the most trusted first.
  • Find who to review. GET /v1/objects?q=trust_score < 1. See Search and conditions.
  • Get told when trust drops. A rule with "when": "trust_score < 1" fires once when a user falls below the line.

Combine with other signals ​

Conditions can combine signals, such as is_scammer >= 60 and trust_score < 1, so an alert only fires when a suspicious message comes from someone without a track record.

Next steps ​