Confidence scoring

Confidence scores indicate the reliability of an automated match on a scale of 0-100. Higher scores mean greater certainty that two transactions represent the same financial event.

Overview

When Matcher identifies a potential match, it assigns a confidence score based on multiple factors. This score defines how the match is handled:

High scores (90+) are auto-approved
Mid-range (60 – 89) scores require review
Low scores (<60) are treated as exceptions

Matcher Confidence Scoring — Matcher confidence scoring.

Score components

Matcher uses a binary weighted scoring system with four components. Each component evaluates to either a full match (1.0) or no match (0.0) — there are no partial scores within a component.

Component	Weight	Points (match / no match)
Amount match	40%	40 / 0
Currency match	30%	30 / 0
Date proximity	20%	20 / 0
Reference match	10%	10 / 0

Amount match (40%)

The amount component has the highest weight because amount discrepancies often indicate different transactions.

Condition	Score
Amounts match (within rule tolerance)	40 points
Amounts do not match	0 points

Amount matching depends on the active rule type. An EXACT rule requires identical amounts; a TOLERANCE rule allows variance within the configured percentTolerance and absTolerance.

Currency match (30%)

Currency verification is binary — currencies either match or they don’t.

Condition	Score
Same currency	30 points
Different currency	0 points

Date proximity (20%)

Date scoring checks whether the transaction dates fall within the configured window.

Condition	Score
Dates within allowed window	20 points
Dates outside allowed window	0 points

The allowed window depends on the rule: an EXACT rule requires the same date (respecting datePrecision), while a DATE_LAG rule accepts a day difference inside its [minDays, maxDays] band (the inclusive flag controls whether maxDays itself counts).

Reference match (10%)

For EXACT and TOLERANCE rules, reference comparison is binary.

Condition	Score
References match (exact or case-insensitive)	10 points
Both references absent	10 points (unless the rule sets `referenceMustSet`, which scores any absent reference as 0)
References do not match, or only one is absent	0 points

FUZZY rules score the reference on a continuous scale. For a FUZZY rule, the 10% reference slot carries a graded ReferenceScore between 0.0 and 1.0 (a similarity measure) rather than a strict 0/1. A near-identical reference contributes close to the full 10 points while one that barely clears the fuzzy gate contributes proportionally less. As a result, FUZZY matches can produce non-multiple-of-10 scores such as 97 or 99. See Possible scores below.

DATE_LAG rules do not score references. For DATE_LAG rules the ReferenceScore is always 0.0, so the 10% reference component contributes 0 points regardless of the reference values. The maximum achievable DATE_LAG score is therefore 90 (40 + 30 + 20 + 0), which by design keeps date-lag matches in the manual-review path.

Calculation formula

The confidence score formula:

confidence = round(
  (amountMatch × 0.40 +
   currencyMatch × 0.30 +
   dateMatch × 0.20 +
   referenceScore × 0.10) × 100
)

Where each factor is either 1.0 (match) or 0.0 (no match). The weights are hardcoded constants and are not configurable per context.

Possible scores

For EXACT, TOLERANCE, and DATE_LAG rules, every component is binary, so the confidence score is always one of these values: 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100 For these rule types, intermediate values (e.g. 87, 72, 55) never occur. (For DATE_LAG the reference component is always 0, so its scores never include the final 10 points.) For FUZZY rules this does not hold. Because the reference component is a continuous ReferenceScore (0.0–1.0), FUZZY matches can produce intermediate scores such as 97 or 99. Note that regardless of the resulting score, a FUZZY match is never auto-confirmed — see FUZZY matches never auto-confirm.

Calculation examples

Exact match (score: 100)

Two transactions with identical values on the same day:

Component	Comparison	Points
Amount	$1,000.00 vs$ 1,000.00 ✓	40
Currency	USD vs USD ✓	30
Date	2024-01-15 vs 2024-01-15 ✓	20
Reference	PAY-001 vs PAY-001 ✓	10

Final Score: 100 → Auto-confirmed

High confidence match (score: 90)

All fields match except reference:

Component	Comparison	Points
Amount	$1,000.00 vs$ 1,000.00 ✓	40
Currency	USD vs USD ✓	30
Date	2024-01-15 vs 2024-01-15 ✓	20
Reference	PAY-001 vs — ✗	0

Final Score: 90 → Auto-confirmed

Medium confidence (score: 70)

Amount and currency match, but date and reference do not:

Component	Comparison	Points
Amount	$1,000.00 vs$ 1,000.00 ✓	40
Currency	USD vs USD ✓	30
Date	2024-01-15 vs 2024-01-25 ✗	0
Reference	PAY-001 vs REC-999 ✗	0

Final Score: 70 → Needs review

Low confidence (score: 40)

Only amount matches:

Component	Comparison	Points
Amount	$1,000.00 vs$ 1,000.00 ✓	40
Currency	USD vs EUR ✗	0
Date	2024-01-15 vs 2024-01-25 ✗	0
Reference	PAY-001 vs REC-999 ✗	0

Final Score: 40 → Exception (below 60)

Confidence tiers

Matcher categorizes matches into tiers based on score:

Tier (score)	System behavior	Typical volume (illustrative)
Auto-Approved (≥ 90)	Automatically confirmed	70-80%
Needs Review (60-89)	Queued for manual review	15-25%
Exception (< 60)	Treated as unmatched	5-10%

How confidence tiers are applied

When Matcher proposes a match, it evaluates the confidence score and applies the following steps:

If the score is 90 or higher, the match is automatically confirmed for EXACT and TOLERANCE rules; FUZZY and DATE_LAG matches always require manual review.
If the score is between 60 and 89, the match is queued for manual review.
If the score is below 60, no match is created and the transaction becomes an exception.
Reviewed matches can be either confirmed or rejected, updating their final status.

This ensures high-confidence matches flow automatically while keeping human control where it matters.

FUZZY matches never auto-confirm

The auto-confirm behavior above applies to EXACT and TOLERANCE rules; FUZZY and DATE_LAG matches always require manual review. Matches produced by FUZZY rules are never auto-confirmed, regardless of their confidence score — even a FUZZY match scoring 90 or above is always queued for manual review. This is by design: a fuzzy reference only contributes the 10% reference slot, so the financial fields (amount + currency + date) alone can already reach the 90 threshold. Capping FUZZY below auto-confirm guarantees that a human reviews the fuzzed reference before the match is committed — “fuzzy proposes, never commits.”

Confidence thresholds

Matcher uses fixed thresholds to determine how matches are handled:

Threshold	Score	Behavior
Auto-confirm	>= 90	Match is automatically confirmed
Match	>= 60	Match is proposed for manual review
Exception	< 60	No match created; transaction becomes an exception

These thresholds are not configurable per context.

Weights

The component weights (40/30/20/10) are hardcoded constants. They cannot be adjusted per context or per rule.

Component	Weight	Rationale
Amount	40%	Amount is the strongest indicator of a valid match
Currency	30%	Currency mismatch usually means different transactions
Date	20%	Date proximity adds confidence but allows for settlement delays
Reference	10%	References are helpful but often missing or inconsistent

Best practices

Monitor tier distribution

Track the percentage of transactions in each tier. Unusual shifts may indicate data quality issues or rule misconfiguration.

Use dry-run mode to test rules

Since confidence depends on which rules match, always dry-run rule changes before committing. This prevents unexpected increases in manual review volume.

Review matches near the 60-point boundary

Periodically review matches just above the exception threshold. These often reveal opportunities for rule improvements.

Understand scores as rule outcomes

For EXACT, TOLERANCE, and DATE_LAG rules, scoring is binary, so a score of 70 means exactly “amount + currency matched, date + reference did not.” Use this to diagnose matching issues. FUZZY rules are the exception: their graded reference component can yield intermediate scores (e.g. 97), so a FUZZY score should be read as “financial fields matched plus a partial reference similarity.”

Overview

Guides

Matcher UI

MCP Server

Operations

Confidence scoring

Overview

Score components

Amount match (40%)

Currency match (30%)

Date proximity (20%)

Reference match (10%)

Calculation formula

Possible scores

Calculation examples

Exact match (score: 100)

High confidence match (score: 90)

Medium confidence (score: 70)

Low confidence (score: 40)

Confidence tiers

How confidence tiers are applied

FUZZY matches never auto-confirm

Confidence thresholds

Weights

Best practices

Next steps

Match Rules

Multi-Currency

​Overview

​Score components

​Amount match (40%)

​Currency match (30%)

​Date proximity (20%)

​Reference match (10%)

​Calculation formula

​Possible scores

​Calculation examples

​Exact match (score: 100)

​High confidence match (score: 90)

​Medium confidence (score: 70)

​Low confidence (score: 40)

​Confidence tiers

​How confidence tiers are applied

​FUZZY matches never auto-confirm

​Confidence thresholds

​Weights

​Best practices

​Next steps

Match Rules

Multi-Currency

Overview

Score components

Amount match (40%)

Currency match (30%)

Date proximity (20%)

Reference match (10%)

Calculation formula

Possible scores

Calculation examples

Exact match (score: 100)

High confidence match (score: 90)

Medium confidence (score: 70)

Low confidence (score: 40)

Confidence tiers

How confidence tiers are applied

FUZZY matches never auto-confirm

Confidence thresholds

Weights

Best practices

Next steps