Certificates of Analysis and Mill Test Reports for Quality Document Extraction

Why quality certificates need their own extraction data

Quality certificates are verification documents, so the model must read a table and then judge it against a limit, which invoice or form data does not teach. An invoice line item has a quantity and a price; a CoA row has a parameter such as assay, water content by Karl Fischer, or heavy metals, a result that may be "<0.5 ppm" or "conforms", a unit, and a limit expressed as a range, a one-sided bound or a pharmacopeial reference. A mill test report (MTR) adds heat numbers, chemistry columns (C, Mn, P, S, Si, Cr, Ni, Mo), mechanical properties (yield, tensile, elongation, hardness), and heat-treatment notes in dense multi-row headers.

Generic benchmarks do not close the gap. FUNSD, for example, has 199 noisy scanned forms labeled for entity linking, which is useful for layout but contains no specification semantics [2]. For table structure and reading order, see our guide to document layout analysis datasets; for header-and-value pairs, see key-value extraction labels. Trade paperwork such as certificates of origin is covered separately in trade document extraction data.

What the regulatory and standards context means for labels

The document type determines what a correct extraction even means, so record it as a label. In metals, EN 10204 is the standard that defines inspection document types, commonly cited as 2.1, 2.2, 3.1 and 3.2, which differ in whether results are included, whether they are specific to the delivered product, and who validates them. A model that treats a certificate without lot-specific results as lot-specific evidence has made a substantive error, so the certificate type should be a classified field, with its definitions checked against the current text of EN 10204 before you rely on them.

In US drug manufacturing, the current good manufacturing practice rules in 21 CFR Part 211 govern how incoming components are tested and released, and how far a manufacturer may rely on a supplier's CoA; confirm the exact conditions with your regulatory team. Many regulated receivers therefore run confirmatory tests on a schedule. For training data, those confirmatory in-house results, where they exist, are the most valuable ground truth: they let you measure where supplier certificates and your own lab disagree.

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

The label layers a buyer should specify

Specify labels at the row level, not the document level, because compliance checks fail on single rows. The minimum useful schema has four layers:

• Header fields: supplier, manufacturer (often different from the distributor), material or grade, lot or heat number, manufacture and retest or expiry dates, PO number, quantity, certificate type, signatory.
• Test table: parameter as printed, a normalized parameter code, test method (for example a USP chapter or ASTM method designation), result as printed, parsed numeric value, qualifier (<, >, ND, conforms), and unit.
• Specification layer: the limit from the buyer's own material specification, not only the limit printed on the certificate, since the two often differ.
• Outcome layer: per-row pass/fail computed against the buyer spec, plus the actual receiving-inspection disposition (accepted, rejected, use-as-is, sent to MRB) and any nonconformance report ID.

The outcome layer is what separates this data from OCR data. Pairing certificates with system-of-record entries is the approach described in documents paired with ERP records, and human corrections from a production pipeline are covered in IDP human correction data.

Illustrative example: invented to show structure; it does not describe an available dataset.

{
  "doc_id": "coa-000412",
  "doc_type": "certificate_of_analysis",
  "cert_type_en10204": null,
  "supplier_token": "SUP_0193",
  "material": "Microcrystalline cellulose",
  "lot_token": "LOT_88F2",
  "rows": [
    {"param_printed": "Loss on drying", "param_code": "LOD", "method": "USP <731>",
     "result_printed": "4.1 %", "value": 4.1, "qualifier": null, "unit": "%",
     "buyer_spec": {"max": 5.0}, "row_pass": true, "bbox": [112, 640, 980, 668]},
    {"param_printed": "Total aerobic count", "param_code": "TAMC", "method": "USP <61>",
     "result_printed": "<10 cfu/g", "value": 10, "qualifier": "<", "unit": "cfu/g",
     "buyer_spec": {"max": 1000}, "row_pass": true, "bbox": [112, 702, 980, 730]}
  ],
  "receiving_disposition": "accepted",
  "confirmatory_test_available": true
}

Failure modes to test before you buy

Ask for a sample that contains the hard cases, because clean single-table CoAs inflate accuracy. The recurring failure modes are:

1. Unit and qualifier loss: "<0.5 ppm" extracted as 0.5, or mg/kg confused with %.
2. Spec confusion: the model copies the supplier's printed limit and passes a row that fails the buyer's tighter spec.
3. Multi-heat MTRs: one report covering several heats, where chemistry rows must link to the right heat number.
4. Rotated and merged headers: chemistry columns with element symbols in vertical text, or merged "Mechanical properties" header cells.
5. Scan quality: faxed or re-scanned certificates with stamps and signatures over the table.
6. Template drift: the same supplier changing its CoA layout between years.

Measure accuracy at row and field level, with separate numbers for value, unit and pass/fail; see document extraction evaluation ground truth for how to build a held-out set. ISO/IEC 5259-2 gives measurable data quality characteristics you can write into acceptance criteria, such as completeness and accuracy of labels [1].

Buyer checklist for a CoA and MTR dataset

Use this checklist to scope a request so suppliers can confirm what they actually hold.

Illustrative example: invented to show structure; it does not describe an available dataset.

Supplier names, lot numbers and pricing on attached POs can be confidential to the receiving company and its suppliers, so plan for consistent tokenization that preserves joins (same supplier, same token) rather than blanket redaction.

How SourceX sources quality certificate data

SourceX sources operational datasets from US companies on request; it does not hold certificate data in stock, and a request does not guarantee a match. You describe the documents and labels you need, SourceX looks for US businesses that hold them, and every release is approved by the supplying company. Each dataset is rights-reviewed for ownership and consents and delivered under a license that defines records, uses, term and delivery. Personal details such as names, emails and phone numbers are removed or replaced before delivery, the method is recorded, and a sample is checked, though no method is perfect.

Diligence materials covering source, rights, preparation and allowed use are prepared per dataset, and delivery runs through private, access-controlled workflows only after an executed agreement and supplier approval. Pricing is not published and is agreed per deal. For the broader category of quality records, see manufacturing quality datasets and buyers in manufacturing, or browse the document AI data hub and the AI data hub. To start a scoped request, use the SourceX buyer intake.

Request certificate of analysis extraction data

Describe the certificate types, industries, label layers and outcome data you need, and SourceX will look for US companies that hold matching records. Nothing is contracted until a supplier agrees, and terms are set in a license per deal. Start your request on the SourceX buyers page.