Page 07 / 08 · Quality & Review

Automation does not mean unchecked

Fully automated processing is only valuable if it recognizes when it is uncertain. Two gates hold back uncertain cases, a journal makes every run traceable, and a correction workstation returns reviewed cases to the chain without re-analysis.

Gate 1 · before analysis

Classification gate

If the confidence of the document type falls below the threshold, the subdocument does not go into analysis but into the review list. The split document remains available for preview and reprocessing. The classifier model threshold overrides a global default value; if none is set, the gate is off.

Effect: no analysis cost for doubtful assignments
Gate 2 · before delivery

Extraction gate

If the extraction result falls below the model's quality threshold, it is still saved — but not delivered. The saved result is the working basis for correction; after that, a delivery attempt without new analysis is sufficient. Without a confidence value, the gate does not apply.

Effect: no bad metadata in the target system
Journal — every document has an end state
Batch
chevron_right
Classified
chevron_right
Analyzed
chevron_right
Mapped
chevron_right
Endpoint:<target>
Succeeded
Skipped with reason
Failed with error text
Classification review
Extraction review

The document key is the classifier output path — it connects all stages of a case. New delivery targets appear automatically in the monitor because endpoint stages are grouped generically.

Correction workstation
label
Override category — reviewer sets the correct document type; the blob metadata is corrected and extraction starts.
edit_note
Correct fields and table cells — changes apply to the saved result, then delivery without re-analysis.
block
Discard — terminal end state in the journal; the document remains in place, nothing is deleted.
Outputs

The extraction result is target- and engine-neutral. What happens to it is decided by adapters — not by the processing code. A new target system adds an adapter; dispatch, storage policy, and monitor remain unchanged.

Document managementObject + properties
REST targets (n per model)Placeholder instead of script
Data exportBatch output
Mandatory targets determine completion; optional errors do not stop the case.
Repeated delivery is keyed idempotently — a retry does not create a duplicate.
Secrets are never stored in the database, only their reference name; an outbound guard checks outgoing addresses.
Known state in branch develop

The write path into document management is currently not active — the connected facade operates in read mode. The REST adapter is therefore currently the only end-to-end working output. The chain itself is unaffected by this: the adapter reports the error as a result, and the case remains parked for repeat attempts.