Governed reconciliation: matching tiers, reason codes, and evidence
By Equipo Quantum Developers
Summarize:
Reconciliation is governable only when each match can be reproduced from versioned rules, reason codes, and evidence, while ambiguous decisions end in separate approval. When a system answers “AI match” but cannot show compared fields, tolerance, version, and accountable reviewer, it has automated a conclusion rather than a control.
Reconciliation is not the search for something similar
Reconciliation establishes whether records from different sources represent the same obligation or movement and which difference remains unresolved. The result may authorize accounting closure, release cases, or initiate investigation. The architecture should therefore minimize false positives before maximizing automatic coverage.
One score blends different problems: missing reference, out-of-policy amount, shifted date, split transaction, duplicate, or conflicting evidence. An ordered set of tiers is safer. Each tier knows what it may decide, which reason it produces, and when it must stop.
Four matching tiers with increasing authority
Tier 1: exact deterministic match
Compare stable identifiers, currency, amount, and expected relationship. A valid unique key plus a versioned rule can close automatically when policy allows. Preserve hashes or references to source records, not only the result.
Tier 2: governed tolerances
Allow explicit date, amount, or grouping differences only through policy. Tolerance belongs to the transaction type, not the model. Record observed value, threshold, currency, time zone, and rule version. Outside the limit, do not “pull” the case closer; move to another tier.
Tier 3: contextual assistance
Use search or AI to propose candidates when the reference is noisy or embedded in text. The system can explain signals—normalized text, counterparty, window, permitted history—but cannot turn ambiguity into certainty. Require adequate separation between candidates and route the remainder to review.
Tier 4: human resolution
The reviewer receives candidates, permitted documents, differences, rule, and available actions. The person resolving a case should not also change tolerances or approve a material exception. Closure preserves identity, time, and reason.
These tiers are not an “intelligence” scale. They separate authority: exact rules resolve what can be proven, tolerances apply policy, AI organizes ambiguity, and people decide irreversible cases.
Reason codes that outlive free text
Define stable result and exception codes:
- MATCH_EXACT_REFERENCE_AMOUNT: exact reference, currency, and amount.
- MATCH_POLICY_TOLERANCE: difference inside an identified tolerance.
- MATCH_GROUPED_COMPONENTS: multiple entries explain one transaction under a rule.
- REVIEW_MULTIPLE_CANDIDATES: no dominant candidate exists.
- REVIEW_MISSING_REFERENCE: a required identifier is absent.
- BLOCK_AMOUNT_OR_CURRENCY: difference exceeds policy.
- BLOCK_POSSIBLE_DUPLICATE: the object appears to be reconciled already.
- UNMATCHED_SOURCE_MISSING: no counterpart exists when the window closes.
When a payment network uses ISO 20022 messages, align internal codes with external states without blindly copying them. The ISO 20022 registration authority publishes external code sets in machine-readable formats and updates them periodically. Store the catalog version and date in use; a valid code may evolve without a message-schema change.
The evidence pack for each decision
Every match should create a compact, reproducible envelope:
| Field | Evidence |
|---|---|
| reconciliation_id | stable, idempotent identity |
| source_record_refs | references and versions on both sides |
| business_object | affected payment, settlement, invoice, or movement |
| tier and rule_version | executed level and exact policy |
| compared_fields | normalized values and differences |
| reason_code | structured explanation of the result |
| candidate_set | considered candidates when applicable |
| actor_or_approver | separated system, operator, and approver |
| decided_at | zoned time and event order |
| outcome | matched, pending, rejected, or reversed |
NIST control AU-3 in SP 800-53 Rev. 5 requires audit records that establish event type, time, location, source, outcome, and associated identities. It is not an accounting reconciliation standard, but it supplies a useful sufficiency test: a reviewer should reconstruct what occurred without relying on an informal screenshot.
Illustrative example: three results, three reasons
This example is strictly illustrative and contains no real data. The fictional policy permits a maximum USD 0.50 difference for one fee class and requires human approval for grouped matches.
- Record A and its counterpart share reference, currency, and USD 125.00. Tier 1: MATCH_EXACT_REFERENCE_AMOUNT.
- Record B is USD 80.00 and its counterpart USD 79.75, with an allowed fee type. Tier 2: MATCH_POLICY_TOLERANCE; the pack stores the USD 0.25 difference and policy version.
- Record C for USD 300.00 has two candidates of USD 100.00 and USD 200.00. Tier 3 proposes grouping; Tier 4 requires a reviewer to confirm relationship and evidence before MATCH_GROUPED_COMPONENTS.
The system never “learns” a new tolerance silently from that approval. The decision creates a separate change proposal with a test and approver.
Operating control and separation of duties
The GAO Green Book organizes an effective internal-control system around operations, reporting, and compliance objectives, with preventive and detective activities. Applied here, configuration, operation, and approval remain separate; aged exceptions are reviewed; control totals are reconciled; and proportionate evidence is retained.
Daily controls should include count and value by state, unprocessed inputs, retries, blocked duplicates, and queue age. Treat tolerance changes as policy changes, not technical tuning. Growth in Tier 3 matches may indicate deteriorating references rather than a better model.
Observable implementation in Quantum
Quantum Automation Center can represent each reconciliation as a business object, follow execution through a timeline, retain references to permitted artifacts and logs, show its exception, and apply controlled permissions or human approval. Analytics can separate coverage by tier, block code, and resolution age. See payment-method reconciliation and the Quantum ontology.
Do not expose full documents on a dashboard when protected references are enough. Traceability does not require multiplying sensitive data.
Counterargument: one model looks simpler
It may be simpler to demonstrate and may find relationships rules did not anticipate. But it moves complexity into evaluation, explanation, and correction. An automatically reconciled false positive is often more expensive to investigate than a case routed to review from the start.
Use the model in Tier 3, where flexibility ranks candidates. Promote it only inside specific boundaries evaluated by case type, while retaining deterministic rules for keys and tolerances.
When not to use AI or probabilistic matching
Do not use it when reliable identifiers and deterministic rules resolve the case. Do not use it without a reference truth for evaluation, with sources lacking sufficient rights or quality, or without people to resolve exceptions. In those settings, the model adds a score, not knowledge.
If volume is low and each case requires expert judgment, an evidence-rich worklist may be better. Automate ingestion, normalization, and preparation while preserving human decision authority.
The test of governed reconciliation
Select one result from every tier. Re-run it with the same inputs, rule version, and code catalog. Verify the same reason appears, then reconstruct who approved and what changed on the object. If the outcome cannot be reproduced, do not raise the automatic threshold. Repair evidence and authority separation first.
Sources
Article topics
