The engine · Deterministic

The deterministic AI engine behind every correction

A seven-step pipeline across 24 endpoints at sub-50ms P95 — every correction reproducible and traceable to a cited regulatory clause, a guarantee LLM-based parsers cannot make.

The ARS AI Engine converts free-text, hybrid or structured payment addresses into ISO 20022-compliant output in a single call, running a deterministic seven-step pipeline across 24 endpoints at sub-50ms P95 latency. Every correction is reproducible and traceable to a cited regulatory clause.

The seven-step AutoCorrect pipeline

Behind every AutoCorrect call, the engine runs the same pipeline end to end, inside one API call.

1

Classify

Detect whether the input is unstructured, hybrid or structured.

2

Validate

Check field presence, character limits and occurrence rules per output mode.

3

Postal-verify

Match against real postal databases across 246 countries with a 0.0–1.0 confidence score.

4

Extract entities

Pull BIC, LEI and IBAN out of misplaced address lines into their correct structured fields.

5

Convert format

Promote town and country into structured fields; remove duplications; convert unstructured → hybrid or fully structured.

6

Apply scheme rules

The regulation-aware step: eight layers of rules applied automatically from three parameters (scheme, party role, execution datetime).

7

Render

Write corrected addresses back into the correct XPaths, preserve all namespaces, and return a complete, XSD-valid ISO 20022 message.

No regulation to encode
You do not encode EU Regulation 2023/1113, EPC153-22, PMPG v1.11 or the CBPR+ XSD. The engine reads them on your behalf.

AutoCorrect vs PreCheck

AutoCorrect
The resolve-and-fix endpoint: submit a message, receive it corrected. One call solves the problem.
PreCheck
A sub-100ms structural validation for keystroke-level UI feedback, or as a cost-optimisation pre-filter before AutoCorrect.

Every decision traceable

The same input always produces the same output, every time. Every decision is traceable through a rule_source field on each of 30 structured reason codes.

Audit-grade provenance
In an audit, rule_source: EPC153-22 §4 points the examiner straight to the regulation — far more defensible than "unstructured address found." This determinism is the substrate that distinguishes ARS from probabilistic, LLM-based parsers.
30 reason codes · six families
FMT — FormatFLD — FieldREG — Regulatory ENT — EntityVRF — VerifyCNV — Convert
Four response statuses
VALIDATEDREPAIREDPARTIALLY_REPAIREDNOT_REPAIRED

Four output modes, bidirectional

A single engine renders to four configurable modes and converts in both directions — OUTBOUND to upgrade for the network, INBOUND to down-render for legacy sanctions or RTGS gateways, even mixed within a single call.

7×70
COMPLIANCE
2×35
MT_LEGACY
3×35
CBPR_PLUS
2×70
EPC

Specifications

MetricSpecification
Latency< 50ms P95 for the full seven-step pipeline
Throughput1,000+ TPS sustained
Capacity10B+ transactions per year
BatchUp to 1,000 transaction sets, or 100 messages, per call
Coverage

246 countries (including 41 SEPA: 30 EEA + 11 non-EEA), 11+ payment schemes (SCT, SCT Inst, SDD Core, SDD B2B, OCT Inst, CBPR+, CHAPS, T2, EURO1, Fedwire and more), and 25+ ISO 20022 message types across pain, pacs and camt.

Frequently asked questions

Yes. A single REST endpoint accepts free-text, hybrid or structured input and returns corrected ISO 20022 output. MCP, IBM MQ, Kafka, SFTP and JDBC channels are also supported.

Sub-50ms P95 for the full seven-step pipeline; the structural PreCheck path is sub-100ms.

Deterministic. Identical input always produces identical output, with every correction tied to a cited regulatory rule_source — a requirement LLM parsers cannot meet.

Yes. The engine writes corrections back into the correct XPaths, preserves all namespaces, leaves untouched elements intact, and returns a complete, XSD-valid ISO 20022 message — no XPath logic on your side.