概述

verify-claim

Verify factual claims against live data before using them in responses, calculations, or decisions. Returns structured verdicts with confidence scores and current truth values.

When to Activate

Use this skill whenever:

A user states a number as fact ("the exchange rate is 0.92", "Bitcoin is at $60K")
You are about to use training-data numbers in calculations or recommendations
A user asks "is it true that..." or "can you verify..."
You need a current price, rate, date, or statistic for an accurate response
You are writing a report or analysis that cites specific figures
A previous response included a number that may have changed since your training cutoff

Do NOT use for: opinions, predictions, subjective claims, private/internal data, or questions about future events.

Workflow

Step 1: Extract the claim

Identify the specific factual assertion. Be precise — "Bitcoin price is above $60,000" is better than "Bitcoin is expensive."

Step 2: Check via MCP tool (preferred)

If @agentutil/verify-mcp is available as an MCP tool, use it directly:

verify_claim({ claim: "The USD to EUR exchange rate is 0.92" })

Step 3: Fall back to HTTP

If no MCP tool is available, call the API:

curl -X POST https://verify.agentutil.net/v1/verify \
  -H "Content-Type: application/json" \
  -d '{"claim": "The USD to EUR exchange rate is 0.92"}'

Optional: add "category" for faster routing — financial, entity, geo, or factcheck.

Step 4: Interpret the verdict

Verdict	Confidence	What to do
---------	------------	------------
`confirmed`	> 0.9	Safe to use. Cite `current_truth` as the verified value.
`confirmed`	0.7-0.9	Likely correct but note the uncertainty to the user.
`stale`	any	The claim WAS true but data has changed. Use `current_truth` instead. Tell the user: "That was previously correct, but the current value is X."
`disputed`	any	Sources disagree. Present both sides — do NOT pick one. Say: "Sources conflict on this — [source A] says X while [source B] says Y."
`false`	> 0.8	Contradicts current data. Correct it: "Actually, the current value is X."
`false`	< 0.8	Likely wrong but not certain. Flag it: "This doesn't match current data (X), but confidence is moderate."
`unknown`	any	Could not verify. Say so honestly — do NOT guess.

Step 5: Act on the result

In calculations: always substitute current_truth for the original claim value
In reports: cite the verified figure with freshness indicator ("as of [freshness]")
In conversations: correct stale data naturally without being preachy
If unknown: proceed with the original value but caveat that it's unverified

Response Format

{
  "verdict": "confirmed",
  "confidence": 0.95,
  "current_truth": "0.921",
  "freshness": "live",
  "source_count": 2,
  "cached": false,
  "request_id": "abc-123",
  "service": "https://verify.agentutil.net"
}

Trending Claims

See what others are verifying:

curl https://verify.agentutil.net/v1/trending

Pricing

Free tier: 25 queries/day, no authentication required
Paid tier: unlimited queries via x402 protocol (USDC on Base), $0.004/query

Agent Discovery

MCP server: @agentutil/verify-mcp (npm)
Agent Card: https://verify.agentutil.net/.well-known/agent.json
Service metadata: https://verify.agentutil.net/.well-known/agent-service.json

Privacy

No authentication required for free tier. Query content is not stored beyond transient cache (max 1 hour). No personal data collected. Rate limiting uses IP hashing only.

版本历史

共 2 个版本

v2.0.1 当前

2026-03-29 12:16 安全安全
v1.0.1

2026-03-11 10:55

安全检测

腾讯云安全 (Keen)

安全，无风险

查看报告

腾讯云安全 (Sanbu)