Confidence Scoring

Understanding Confidence Scores

Confidence scores measure how certain CargoLint’s AI is about extracted field values. Scores range from 0.0 (no confidence) to 1.0 (absolute confidence).

Each extracted field receives an individual confidence score based on:

• Document quality - Clarity, resolution, legibility
• Field location - Position and context in document
• Text patterns - Matching known formats and standards
• Field completeness - Availability of supporting data

Per-Field Scoring

Every extracted field includes a confidence score:

{
  "extracted_fields": {
    "invoice_number": {
      "value": "INV-2026-001",
      "confidence": 0.98
    },
    "invoice_date": {
      "value": "2026-03-02",
      "confidence": 0.94
    },
    "total_amount": {
      "value": 5250.00,
      "confidence": 0.87,
      "currency": "USD"
    },
    "buyer_phone": {
      "value": "+1-555-0123",
      "confidence": 0.62
    }
  }
}

Review Threshold

CargoLint uses a 78% (0.78) overall confidence threshold to determine whether a document requires human review. Documents scoring below this threshold are automatically routed to the review queue. The threshold is the same for every document type, so scores are comparable across invoices, packing lists, bills of lading, and certificates of origin.

The overall score is the average of per-field confidences, scaled down multiplicatively by each validation issue:

• Error-level issues (e.g., missing required fields): each reduces the score by 20%
• Warning-level issues (e.g., calculation mismatches, format inconsistencies): each reduces the score by 10%
• Informational notes: no score impact

How Scores Affect Routing

CargoLint automatically routes documents based on confidence scores:

Pending Approval Workflow:

1. Overall document confidence >= 0.78
2. Every line item’s HS code confirmed by a reviewer
3. Scan quality above the OCR threshold
4. Document awaits a one-click approval

Review Queue Workflow:

1. Overall document confidence < 0.78, an unconfirmed HS code, or a degraded scan
2. Document flagged as requires_review
3. Reviewer confirms or corrects each field
4. Validated data saved

HS Code Confidence

HS code suggestions have their own thresholds, separate from the document score:

• >= 0.70 - The top candidate is shown as a recommendation to confirm
• 0.25 - 0.70 - Suggestions are shown; the line is flagged as missing a code
• < 0.25 - Suggestions below this floor are dropped; manual classification needed

Suggestions are always suggestions: CargoLint never applies an HS code automatically, and a document with an unconfirmed code always goes through review.

Declared codes are validated at the 6-digit (WCO) level. A longer national code (HTS-10, TARIC-10) is assessed on its first 6 digits, and the reviewer confirms the national suffix - the flags say so explicitly when this applies.

Improving Scores Over Time

User Corrections

When you correct an extracted value in the review queue, CargoLint logs the correction. HS code corrections feed back into your organization’s future suggestions automatically; extraction corrections accumulate as training data used to evaluate and improve extraction models.

A document’s confidence score describes extraction quality at processing time. After you complete review, CargoLint shows the document as Reviewed instead of the original score.

Interpreting Scores by Field Type

Confidence thresholds vary by field category:

Structured Fields (Numbers, Codes)

• Invoice numbers: Typically 0.90+
• Dates: Typically 0.92+
• HS codes: Typically 0.85+
• Currency amounts: Typically 0.88+

Lower confidence indicates: Handwritten data, unusual formatting, or poor document quality.

Alphanumeric Fields (Names, Addresses)

• Shipper/consignee names: Typically 0.80+
• Addresses: Typically 0.75+

Lower confidence indicates: Handwritten names, non-Latin characters, or abbreviated text.

Unstructured Fields (Descriptions, Notes)

• Line item descriptions: Typically 0.70+
• Marks and numbers: Typically 0.65+

Lower confidence indicates: Abbreviations, technical jargon, or non-standard formatting.

Best Practices

• Start conservative with auto-import thresholds, relax as confidence improves
• Monitor low-confidence fields to identify document quality issues
• Use confidence scores to prioritize manual review queue
• Track score trends over time to spot document quality or template drift