What it does

content-trust-check runs the checks an agent typically wants before publishing or trusting a block of text: is it AI-generated, is it safe, and is it a copy of something else. 

{
  "text": "The x402 protocol enables HTTP-native payments between autonomous agents using stablecoins."
}

returns each signal plus a rolled-up score: 

{
  "ai_generated": { "likelihood": 0.18, "label": "likely_human" },
  "moderation": { "flags": [], "safe": true },
  "originality": { "simhash": "0x1a2b3c4d5e6f7081" },
  "verdict": { "trust_score": 100, "notes": [] }
}

Pass reference_url and it also scrapes that page, hashes it, and reports the Hamming distance between the two SimHashes as originality.reference_distance with a near_duplicate flag.

Why

AI detection, moderation, and near-duplicate checks are three separate paid calls in the catalog today (ai-content-detector, moderate-content, content-simhash). content-trust-check runs all three in parallel behind one call and adds a deterministic roll-up score, which is the piece none of the three individually provide.

How the score works

verdict.trust_score is computed in plain code, not by an LLM: it starts at 100 and subtracts a fixed penalty per risk signal a leg actually returned (higher AI-generation likelihood, a moderation block or review verdict, or a near-duplicate reference match), floored at 0. A leg that fails or is skipped contributes no penalty and is called out in verdict.notes instead of being assumed clean.

Degradation

At least one of the three core legs (ai-content-detector, moderate-content, content-simhash) must succeed for the call to return; the others degrade individually. The reference-URL comparison is always optional.

Price: $0.05.