Skip to main content
Every time an operation writes a value to the catalogue using a model, it attaches a confidence score (0–100) and citations to that value. The score governs whether the write is auto-applied, queued for review, or discarded — and it travels with the value as part of its provenance.

​
How confidence is computed

Claro’s confidence score combines four factors:
FactorWeightWhat it measures
Source reliability40%How trustworthy is the underlying source? First-party documents and authoritative sites score higher than unknown web sources. Knowledge Base content scores higher than open-web search.
Content consistency30%Do multiple sources agree? Consistent cross-source signals raise the score; conflicting signals lower it. Single-source answers receive moderate confidence.
Model certainty20%How confident was the model in its own output? Low variance in the model’s internal probability distribution indicates higher certainty.
Retrieval quality10%How relevant was the retrieved context to the task? Strong semantic match and comprehensive coverage push this higher.

​
Confidence bands and their meaning

BandScoreWhat happens
High85–100Auto-apply (if at or above the auto-apply threshold).
Medium60–84Typically queued for review.
Low0–59Queued for review or discarded, depending on the reject threshold.
Thresholds are configurable per operation and per attribute. Defaults are conservative — most teams raise them incrementally as they gain confidence in an operation’s output quality.

​
Three thresholds per operation

SettingEffect
Auto-apply thresholdConfidence at or above this → the value is written to the record without review.
Review thresholdBetween auto-apply and this → the value is queued in Notifications for human review.
Reject thresholdBelow this → the proposed value is discarded with a logged reason.
Set these under the catalogue’s Config tab. You can override per attribute (e.g. a stricter threshold for pricing fields than for descriptions).

​
Citations

Every auto-applied or queued value carries citations — links to the specific sources the model drew from. Citation types:
  • Knowledge Base — document name, page number, and the relevant passage.
  • Web sources — URL and the extracted text snippet.
  • Geolocation — map service and data provider.
  • File extraction — document section and page reference.
  • Intra-catalogue — another record and attribute on the same or related catalogue.
Citations are visible inline on every cell, in the review item in Notifications, and in the operation’s run history. They form part of the value’s provenance and are retained even after auto-apply.

​
Working with confidence in practice

Reviewing outputs:
  • In Notifications, sort review items by confidence ascending to tackle the most uncertain first.
  • For large batches, review the bottom decile before raising the auto-apply threshold.
Improving confidence:
  • Add relevant Knowledge Bases — authoritative content is weighted higher than open-web search.
  • Write tighter prompts that define the exact format and scope of the answer.
  • Provide examples of the desired output.
  • Cross-reference multiple sources: configure enrichment to search more than one source and let Claro compare the results.
When confidence stays low:
  • A persistently low-confidence field usually means the source data doesn’t cover it, or the prompt is ambiguous. Fix the prompt or the source rather than just lowering the threshold.
  • Use the review queue as a training signal: approved values become positive examples that improve future matching in the similarity graph and in human-in-the-loop feedback flows (Dedicated plan).

​
Rollback

Every auto-applied write is reversible. From the record’s per-field history, you can roll back to any prior value. A roll-back is itself a logged event, with your identity as the source.