Six layers. Every quantity defensible.

Vector PDFs (born-vector engineering drawings) are extracted directly. Raster PDFs and scans go through the vision-model OCR pipeline. Both produce structured tokens with sheet + region + page-coordinate metadata captured at this layer.

ingestion_artifact { id, project_id, sheet_id, ingestion_class: 'vector'|'raster', source_uri, raw_extraction_blob, processed_at }

Per-trade-context symbol recognizer matches ingestion tokens against the org's symbol library. Each detection captures a confidence score, the recognizer model version, and the symbol-library entry used. Active-learning pipeline updates symbol entries on confirmed annotations.

detection { id, ingestion_artifact_id, symbol_library_entry_id, confidence, recognizer_model_version, region_polygon, page_coords, created_at }

Detected items link bi-directionally to spec section paragraphs via the spec-link resolver (CSI 40% / keyword 35% / equipment-tag 25%). Each link carries a score + evidence path that survives serialization to the audit log.

spec_link { id, detection_id, spec_section_id, paragraph_offset, score, evidence_signals[], created_at }

Every takeoff line carries a complete evidence chain: detection → spec link → recognizer model version → ingestion artifact. Click any quantity in the UI and you walk the chain to the original sheet + region. Audit log records every chain traversal.

takeoff_line { id, project_id, csi, qty, unit_cost, ext_cost, source_chain: [detection_ids, spec_link_ids, ...], chain_hash }

Append-only audit log captures every chain modification, evidence-trail traversal, recognizer training event, and human override. Hash-chain validation makes tampering detectable. Replicated to write-once object storage with 7-year retention.

audit_event { id, event_type, actor_id, target_object_type, target_object_id, payload_hash, prev_event_hash, immutable_at, retained_until }

Customer-facing exports (proposal PDF, RFQ packets, audit-trail CSV) preserve the evidence chain in human-readable form. Each line in the proposal has a footnote referencing the sheet, region, and recognizer call that produced it.

proposal_artifact { id, project_id, format: 'pdf'|'csv'|'docx', evidence_chain_embedded: bool, generated_at, sha256 }