Blueprint: adaptive document triage

Objective

Triage incoming documents into workflow queues while extracting structured fields, identifying risk, and abstaining when confidence is insufficient. The system evolves by adding specialists for new document families and retiring redundant ones.

Capability packages

• format and language detector;
• document-family classifier;
• field extraction specialists;
• risk and policy classifier;
• confidence calibrator;
• deterministic schema validator;
• human-review queue adapter.

Request plan

FUNCTION triage(document)
    metadata <- DETECT_FORMAT_LANGUAGE_AND_DATA_CLASS(document)
    family <- CLASSIFY_DOCUMENT_FAMILY(document, metadata)

    IF family.confidence < family_threshold
        RETURN HUMAN_REVIEW("unknown family")
    END IF

    extractor <- ROUTE_TO_VERIFIED_EXTRACTOR(family.label, metadata)
    fields <- extractor.EXTRACT(document)
    validation <- VALIDATE_SCHEMA_AND_CROSS_FIELDS(fields)
    risk <- RISK_SPECIALIST_CLASSIFY(document, fields)

    IF NOT validation.pass OR risk.requires_human
        RETURN HUMAN_REVIEW_WITH_EVIDENCE(fields, risk, validation)
    END IF

    RETURN ROUTE_WORKFLOW(fields, risk)
END FUNCTION

Breeding strategy

Cluster human-review cases by failure signature. Create a new specialist only when a stable, sufficiently large niche exists. For small temporary clusters, improve routing, retrieval, or deterministic rules instead.

Candidate operators include adapter training, distillation from a larger extraction model, taxonomy update, confidence recalibration, and quantization for high-volume families.

Evaluation

Use exact field accuracy, critical-field recall, family confusion, calibration, abstention, human-review volume, latency, cost, and downstream workflow error. Maintain temporal holdouts to test new document templates.

Safety and governance

Sensitive documents stay within approved jurisdictions. Models return structured outputs only. Deterministic validators enforce formats and cross-field constraints. High-risk or low-confidence outcomes require human review; the model cannot suppress that requirement.

Population management

Retire specialists when their document family disappears, merge those with indistinguishable behavior, and preserve rollback for taxonomies and extractors. Monitor router starvation and class drift.