90 Million Characters.
No Shortcuts.

Document intake and structural mapping

RECON ingests the document and builds a coordinate-resolved word map — every token assigned a page position, bounding box, and structural element type — before any detection begins.

Document type classification and specialist routing

SENTINEL classifies the document type and activates the appropriate specialist detection configuration, including the domain gate that controls whether L6 SPECTRE fires.

Broad structured identifier scan

VANGUARD runs a broad identifier pass over normalised document text, targeting contact identifiers and government-format numbers present across document types; candidates are recorded for later fusion.

Australian-specific identifier detection

MARSHAL targets Australian government-issued identifiers — TFN, Medicare, ABN, ACN, BSB, and AHPRA — applying country-specific format validation where applicable.

Service credential and access token detection

CIPHER detects service credentials, API tokens, and platform access keys embedded in document text using tuned patterns across common cloud and SaaS formats, with reduced-confidence output flagged for review.

Domain-gated industry-specific entity detection

SPECTRE applies domain-gated detection for industry-vertical entities — coverage activates when SENTINEL classifies the document as matching the relevant domain pack, and does not run outside its configured scope.

Spatial layout and label-proximity analysis

NAVIGATOR uses the coordinate map from RECON to analyse spatial token relationships — financial identifiers adjacent to field labels receive elevated detection confidence; isolated numbers in running prose do not — backed by the engine's 97%+ measured F1 baseline.

Structured form field extraction

ADJUTANT processes structured input elements — form fields, table cells, checkbox groups — extracting values paired with their field-label context as validated detection candidates, distinct from visually similar values in free text.

Context-aware names — narrative prose

OVERWATCH detects person and organisation names in narrative prose using context-aware signals; marked critical — failure triggers pipeline abort.

Context-aware names — boundary-spanning

TEMPEST runs a second pass to catch names spanning line breaks or chunk boundaries that L9 OVERWATCH may fragment; marked critical.

Context validation — false positive elimination

GUARDIAN validates pooled detection candidates from L3–L10 in context to filter false positives — road names, generic job titles, organisational terms — before fusion; marked critical.

Multi-layer detection fusion

NEXUS consolidates detections from all preceding detection layers (L3–L11) into a unified candidate set — overlapping detections deduplicated, false-negative and false-positive corrections applied, with a safety limit of 5,000 mentions sorted by confidence descending.

Confidence scoring and score emission

HERALD scores the fused detection clusters — weighting by zone (header, body, footer, form field), domain profile, and configuration mode — producing the finalClusters set, each carrying a composite confidence score; nothing proceeds to redaction without a quantified confidence.

Context signal arbitration and confidence adjustment

ARBITER applies co-occurrence signals and contextual rules as the final confidence adjustment — boosting label-adjacent values and suppressing footer noise — logging signal match counts for observability; non-critical, pipeline continues if no signals match.

Allowlist enforcement — final validation gate

WARDEN is the last stage before redaction output — applying the configured allowlist to remove any permitted values from the final set; nothing proceeds to redaction without passing every preceding layer and clearing the allowlist gate.