Designing document reconciliation around evidence and decisions

A document reconciliation system should turn imperfect files into reviewable evidence, not quietly turn extracted text into approved business facts. This implementation note describes a practical boundary between extraction, normalization, comparison and decision.

Four layers with different responsibilities

The intake layer stores each source file, sender, arrival time and version. An extraction layer proposes fields with the page or region that supports each value. A normalization layer aligns units, dates and identifiers without deleting the original text. A comparison layer produces explicit exceptions. The decision layer records who resolved an exception, what they decided and which evidence they used.

Keeping these layers separate matters because OCR confidence is not the same as business authority. A clearly read invoice quantity can still be commercially wrong. A low-confidence weight may be correct but require a reviewer to inspect the page.

Field-level provenance

For every compared value, store a document ID, version, field name, raw value, normalized value and extraction status. A case can then say that an invoice contains 120 cartons while a packing list contains 118, and show both images. The system should avoid a single merged quantity until a reviewer has chosen the appropriate interpretation.

State transitions

Useful states include awaiting document, extraction needs review, mismatch detected, awaiting source correction, ready for reviewer and decision recorded. Each transition should have an event and actor. A new document version should reopen affected checks instead of overwriting the earlier exception.

A bounded pilot

Start with a small set of document types and high-value fields. Test missing pages, revised files, different units, duplicates and genuine exceptions. Measure extraction corrections and reviewer effort before increasing automation. These are design principles for a controlled implementation, not a claim of production performance.

Kaliits explains the buyer workflow separately at https://kaliits.com/solutions/import-dossier-control . This note describes the engineering shape behind such a pilot.