Trust

Every run returns a trust report alongside the output. It accounts for the evidence behind each claim in the answer:

Every claim is cited. Each statement in a prose answer (and each value in a structured one) links to the exact URLs it came from, with verbatim excerpts from the page.
Every claim is graded. high, medium, or low confidence, with the reasoning included.
Every source is classified. Sources are typed (official, news, social, academic) and ranked primary or secondary for the claim they support.

Example

A prose answer comes back with numbered callouts:

Datadog anchors the premium end of the market, with Pro pricing starting
at $15 per host per month [1]. Grafana Cloud takes the opposite approach:
a free tier and usage-based pricing aimed at teams that want to start
small [2]...

Each [n] resolves to a claim in the trust report:

{
  "confidence": "high",
  "reasoning": "Pricing claims verified against official vendor pages; positioning corroborated by independent sources.",
  "sources": [
    {
      "url": "https://www.datadoghq.com/pricing/",
      "title": "Datadog Pricing",
      "type": "primary",
      "source_category": "official"
    },
    {
      "url": "https://grafana.com/pricing/",
      "title": "Grafana Cloud Pricing",
      "type": "primary",
      "source_category": "official"
    }
  ],
  "claims": [
    {
      "callout": 1,
      "confidence": "high",
      "reasoning": "Backed by a primary source (official)",
      "citations": [
        {
          "url": "https://www.datadoghq.com/pricing/",
          "title": "Datadog Pricing",
          "excerpts": ["Pro: starting at $15 per host, per month"],
          "source_category": "official"
        }
      ]
    },
    {
      "callout": 2,
      "confidence": "medium",
      "reasoning": "Backed by a secondary source (news)",
      "citations": [
        {
          "url": "https://devopsweekly.example.com/observability-pricing",
          "title": "The state of observability pricing",
          "excerpts": ["Grafana Cloud's free tier remains the entry point for smaller teams"],
          "source_category": "news"
        }
      ]
    }
  ]
}

Each claim carries the value from the answer, the page it came from, and the exact sentence on that page that supports it. Note that claim [2] grades medium — it is supported by a single news source — so the report tells you exactly which claims to double-check before acting. Use it to verify the agent’s work, or pipe the citations into your own review UI.

Claim Trust panel showing confidence, reasoning, and the quoted source for a citation — Click any citation in Ask Nimble to see its confidence, reasoning, and the quoted source behind it.

Confidence rules

Grades follow fixed rules, applied the same way to every claim in every run:

Primary source → high. Any citation from an official source — or a news/social/academic source when that is the kind of evidence the task calls for — grades the claim high.
Independent corroboration → high. With no primary source, secondary citations from two or more different domains also grade high. Two articles from the same outlet do not corroborate each other — independence means different domains.
Single secondary source → medium. Supported, but by one source only.
No usable citation → low. A value with no citation, or one that failed schema validation, is graded low.

The reasoning field states which rule applied: “Backed by a primary source (official)”, “Supported by 3 independent sources”, “Backed by a secondary source (news)”.

Trust for structured outputs

With an output schema, claims are keyed by JSON path instead of callout number. Every value in a generated dataset carries its own evidence:

{
  "path": "$.vendors[0].market_share",
  "confidence": "medium",
  "reasoning": "Backed by a secondary source (aggregator)",
  "citations": [
    { "url": "https://...", "excerpts": ["Datadog holds an estimated ..."] }
  ]
}

Structured trust also covers missing data:

Verified absence. A null value backed by citations grades high — the sources show the data does not exist (a company with no funding history, a product with no public price).
Schema validation. A value that does not match the output schema is graded low, so structural errors are always flagged.

Overall confidence for datasets

A structured run’s overall confidence is computed from two measurements, and takes the lower of the two:

Measurement	What it asks
Fill rate	How many of the cells the schema asked for were filled?
Claim-trust ratio	Of the researched cells, how many graded `high` or `medium`?

The reasoning states which measurement set the grade: “High confidence - cells fill rate is 96% and trust per claim ratio is 91%”. Values supplied in input_data (pre_existing, in enrichment runs) are excluded — only researched values are graded.

Confidence grades

Grade	Meaning
`high`	Corroborated by authoritative sources — safe to act on
`medium`	Supported by weaker or fewer sources — spot-check before relying on it
`low`	Best available signal; treat as a lead, not a fact
`pre_existing`	Supplied in `input_data` — carried over, not researched

The run also carries an overall confidence and reasoning for the answer as a whole.

trust.claims is designed to be machine-consumed: route low-confidence values to human review, accept high ones automatically, and render citations inline in your product.

Next Steps

Quickstart

See a trust report on your own task

Enrichment

How pre_existing works in enrichment runs

Run result schema

The full trust object, field by field

Getting Started

Web Tools

Extract Templates

Web Search Agents

Jobs

Guides

SDKs

Admin

Proxy

Example

Confidence rules

Trust for structured outputs

Overall confidence for datasets

Confidence grades

Next Steps

Quickstart

Enrichment

Run result schema

​Example

​Confidence rules

​Trust for structured outputs

​Overall confidence for datasets

​Confidence grades

​Next Steps

Quickstart

Enrichment

Run result schema

Example

Confidence rules

Trust for structured outputs

Overall confidence for datasets

Confidence grades

Next Steps