Skip to main content
Quanta Meridian logo
Quanta Meridian Builds

Contained Python build

Document and Request Triage

A Finance and Operations intake receives mixed invoices, credit notes, receipts, evidence documents and reporting requests. The build prepares a suggested record and queue, but holds contradictory or incomplete items for a named reviewer.

The decision

Should the intake analyst accept the suggested class and fields, edit them, reject the record or send it to a different specialist queue?

Generated credit note DOC-CRN-000417 followed through extraction, review and correction

Prepared record

Credit note suggested
Supplier
Generated Supplier 053
Credit note
CRN-000417
Date
06 Jan 2026
Displayed gross
£10,946.65
Related invoice
INV-000410
Purchase order
PO-200417
Held for review

Positive credit amount and mixed invoice wording need a person to decide the posting treatment.

  • The credit amount is missing or does not have a negative sign.
  • The document heading deliberately combines two business-record types.
The cyan path prepares the record. The amber branch appears only when a field conflict or blocking rule requires human judgement.

The working situation

One shared intake, five kinds of work.

A multi-site service company uses one shared intake for supplier documents, expense receipts, control evidence and reporting requests. Sending a credit note to the invoice queue or treating an incomplete request as ready creates rework and can hide the reason for a later correction.

The build prepares a record for the intake analyst. It does not approve a payment, accept control evidence or decide what analysis a reporting request needs.

  1. 01Supplier invoicesAccounts payable
  2. 02Credit notesAdjustments
  3. 03Expense receiptsExpenses
  4. 04Control evidenceAssurance review
  5. 05Reporting requestsReporting lead

One record in full

DOC-CRN-000417 is recognised as a credit note, then held for review.

A score of 99.1% is retained as a model output, not presented as a calibrated probability. 2 deterministic checks still require a person.

  1. 01
    ExtractionExtract and anchor the fields

    9 retained fields connect supplier, credit-note, date, amount, invoice and purchase-order values to page one.

  2. 02
    Model suggestionSuggest credit note

    The text classifier selects the credit-note class with a retained score of 99.1%. The score does not release the record.

  3. 03
    Rule checksHold for review

    The positive displayed amount and “invoice adjustment” heading fail 2 release-blocking checks.

  4. 04
    Human correctionRecord the posting adjustment

    The accounts payable analyst retains the displayed value and records -£10,946.65 as the posting adjustment.

  5. 05
    Audit historyAppend the decision

    9 events retain the document, suggestion, rules, correction and final accounts-payable adjustments queue.

How the decision is prepared

The model recognises patterns. Rules protect declared business conditions.

Pattern model

Suggest the record type

A TF-IDF and logistic-regression baseline reads retained text and scores six generated classes.

  • Class suggestion
  • Ranked scores
  • Model and dataset version
Deterministic checks

Test the fields and release conditions

Named patterns extract fields, while rules test required values, dates, amounts and contradictory wording.

  • Page and character anchors
  • Pass or blocking failure
  • Observed value and explanation
Named reviewer

Accept, edit, reject or reroute

The person sees the page, suggestion, extracted fields and failed checks before recording a reasoned decision.

  • Original suggestion retained
  • Correction stored separately
  • Append-only audit history

Evaluation before a pilot

A strong generated test result still leaves real questions unanswered.

Template families are separated before training and test. The page shows the clean classification result beside extraction errors, abstention and two known policy failures.

Generated test classification900 unseen generated documents
Invoice210correct in generated test
Credit note135correct in generated test
Receipt165correct in generated test
Evidence150correct in generated test
Request150correct in generated test
Unknown90correct in generated test

This clean diagonal is a result for deterministic generated templates, not an accuracy claim for real invoices, scans or handwriting.

Retained review policyA suggestion can be correct and still be held.
770 accepted130 held
Generated test coverage
85.6%
Wrong accepted routes
0
Known policy failures
2

The two known failures are invalid-date reporting requests. They remain in the error register and block any operational pilot.

Field extraction98.6%

of generated test documents had every expected field exactly matched.

Calibration decisionNot selected

The validation partitions had no class error, so wrong-prediction calibration remains unobserved.

Operational pilotNot approved

Real documents, privacy, security, access, retention and monitoring have not been tested.

Native reviewer evidence

The reviewer sees why the document stopped.

This screenshot comes from the contained local service. It keeps the generated page, model suggestion, extracted fields, failed checks and human correction in one review path.

Review cases
4
Decisions
4
Audit events
31

Correction and audit

The reviewed answer never overwrites the first suggestion.

01Document receivedSystem record
02Text and fields retainedSystem record
03Credit note suggestedSystem record
04Blocking rules recordedSystem record
05Reviewer edit savedHuman decision
06Adjustments queue assignedHuman decision

What this build does not prove

Generated documents support reproducible testing, not a production claim.

  • All documents are deterministic generated records; the results do not measure performance on real supplier files, scans or handwriting.
  • The retained scores are not calibrated probabilities. Calibration was measured but not selected because no class error occurred in the retained validation partitions.
  • Two invalid-date reporting requests passed the retained review policy and require a stronger chronological rule before any pilot.
  • The local service and reviewer workflow do not prove production access control, retention, monitoring or cloud deployment.
Technical documentationOpen the model, rules, tests, reviewer workflow and rebuild boundary

Runnable build

The fixed-seed generator rebuilds 6,000 documents across 68 template families. The locked Python package includes a command-line interface, local service, reviewer screen and container definition.

Retained checks

Corpus integrity, template separation, class and field evaluation, threshold analysis, audit-chain verification and responsive reviewer-screen checks.

Model card

The model card states the training data, intended use, measured results, score boundary and reasons an operational pilot is not approved.