How to read a mill test certificate without losing the evidence

First pass: map the document

Find the document title, certificate number, issue date, revision and page count. Locate the organizations named on the document and distinguish their stated roles: manufacturer, distributor, customer or inspection organization. A prominent logo or address does not, by itself, identify the manufacturing mill.

Scan the page structure before transcribing numbers. Note which tables contain product rows, heat references, grade designations, chemical results or mechanical results. Check whether legends and notes continue onto a second page. Different issuers use different layouts; a field’s position is not a reliable definition of its meaning.

SSAB’s Toolox booklet includes a manufacturer-specific certificate example and reading guide on printed pages 6–7. It separates administrative references, product information and test results. Use it to practise finding labels, not as a universal template for every mill.

Second pass: pair labels, values and rows

For each value you need, capture the printed label, the value and the page. If a table heading applies to several rows, keep enough context to identify the intended row. Preserve punctuation and leading zeros in identifiers. An OCR result such as O08419 may need a visual check against 008419.

Treat grade designations as exact document claims. Similar-looking names are not a basis for deciding that two grades are interchangeable. Review test results against the purchase requirements and applicable specifications with a qualified person; MeltLedger’s heat-and-grade comparison is not a chemical or mechanical conformity assessment.

Separate evidence from interpretation

A useful note says “page 2 states Heat H-008419,” rather than simply “correct.” Record what you checked and what remains unknown. If the document states a manufacturing country but no melt-and-pour country, keep those as separate facts. A missing statement should remain missing.

When the PDF is a scan, inspect the original image after using AI. When it contains selectable text, still check the displayed page: text extraction can lose column structure. Mark a field reviewed only after the human check, then keep the original and the exported review together.

Evaluate an AI mill certificate reader against a checked reference

Before using an AI reader, make a small reference set of documents you are authorized to process. Include the kinds you actually receive: a clear digital PDF, a readable scan and, if relevant, a document with several heat rows or a later revision. Read each original yourself first. A convenient clean sample alone tells you little about the harder documents in your inbox.

In a worksheet, record the certificate and revision, field label, exact value, source page and relevant row. Separate fields that are absent from fields that are present but unreadable. This is your checked reference for comparing suggestions; it is not a certificate of authenticity or a standardized accuracy benchmark.

• Check the value and its context together. The right heat number attached to the wrong shipment row is still a wrong result.
• Count unsupported suggestions separately from omissions. A guessed country is different from a blank field that needs investigation.
• Check whether the cited page actually supports the claimed field. A matching word elsewhere on the page is insufficient.
• Record corrections and total review time, including opening source pages and resolving exceptions. Do not measure only the time until a summary appears.

Record exceptions before deciding whether the reader helps

For each required field in each document, note whether the first suggestion agrees with your checked reference, needs correction, misses visible evidence, or remains unresolved because the source is unclear. Keep a separate note for a suggestion that supplies a value the source does not state. These are evaluation notes for your own worksheet, not additional automatic status labels in MeltLedger.

Report counts with their context: which documents, layouts and fields you inspected, and how many remained unresolved. A small trial can expose a workflow problem; it cannot establish an accuracy percentage for every supplier or prove compliance. If the source itself is unreadable, obtain clearer evidence before treating your reference value as settled.

What can be automated, and what still needs a decision?

MeltLedger’s optional AI proposes five evidence fields: certificate identifier, manufacturing mill, heat number, steel grade and an explicitly stated melt-and-pour country. It helps prepare a review; it does not validate chemical or mechanical test results against a specification, authenticate the issuer, or approve a shipment. You still check the original page and confirm the field yourself.

Keep extraction, reference comparison and acceptance as separate steps. Comparing a heat or grade with a reference you entered answers whether those recorded values agree. Acceptance may also depend on the order, inspection requirements and checks outside the PDF. LRQA’s Type 3.2 service describes a specialist inspection process; reading a document with AI does not perform that process.

A repeatable reading sequence in MeltLedger

Open a PDF or the fictional sample, review the five evidence fields, and use optional AI only when you want suggestions. Follow source-page references, correct mistakes and record notes before saving. Compare manually entered shipment heat and grade references if you have them.

Start with a single certificate whose content you can independently verify. The free trial includes five unique certificates, so you can check whether this workflow fits your documents before subscribing.

Fictional AI-reader evaluation notes

Common questions