Mill test reports: how to read one, and where automated reading fails

What a mill test report is

A mill test report, also called a mill cert, an MTR, an MTC or a CMTR, is the mill's quality document for one heat of steel. It records what the mill measured on samples from that heat: the chemical analysis and the mechanical properties. Everything downstream, the tag on the bundle, the bill of lading, your receiving record, the fabricator's traceability file, points back at that heat number and that report.

Two other documents get confused with it. A certificate of conformance is a statement that material meets a specification, and on its own it carries no test data. A transfer cert is a distributor passing the mill's report through, often retyped onto the distributor's letterhead with the mill's original customer masked. That is a normal commercial practice with one rule attached: the distributor may add a conformance statement, and may never alter the mill's test data. The retype is where both honest typos and dishonest edits enter, which is why an experienced receiver asks for the original mill report whenever a transfer cert looks off.

Export-oriented certs often carry EN 10204 "3.1" language. A 3.1 certificate is issued by the mill's own authorized inspector, independent of the production department. It is still the mill vouching for the mill, which is fine, and is not the same as third-party inspection.

The eight blocks on the page

Layouts vary by mill, but the same eight things are always there. If one is missing, that is the first finding.

1. Mill identification. Name, plant location, usually a document number.
2. Customer and order block. The mill's customer, which is usually the service center rather than the end user, the mill order number and sometimes the customer's purchase order.
3. Product description. Section or size, length, and the specifications certified, with edition years. Mills commonly certify one heat to every specification its chemistry and mechanicals satisfy, so A992, A572 Grade 50 and A36 on a single wide-flange line is normal.
4. Heat number. The key that joins this page to the tag, the bill of lading and your records. Formats are mill-specific.
5. Chemical analysis. One row of element percentages for the heat: carbon, manganese, phosphorus, sulfur, silicon, copper, nickel, chromium, molybdenum, vanadium, columbium, and often a computed carbon equivalent. Mills print two decimal places for most elements and three for the low residuals, phosphorus and sulfur.
6. Mechanical properties. Yield, tensile, elongation with its gauge length stated, and, when ordered, bend, hardness, or Charpy impact values with the test temperature and specimen orientation.
7. Country of melt and manufacture. The melted-and-manufactured statement, which is load-bearing on Buy America work and binary: there is no partial credit.
8. Certification statement and signature. A quality manager or metallurgist, dated.

How to check one in five minutes

This is the check a good receiver runs at the dock. None of it needs a metallurgist, and none of it needs software. It needs the grade table for the specification on the cert and a habit.

1. Three heat numbers, one value. The heat on the bundle tag, the heat on the cert, and the heat on the bill of lading must match. On certified work, material with no cert on file for its heat does not enter the job. "Certs to follow" is a state a shipment can be in, not an override.
2. Grade and product form agree. Pipe is A53 or A106, hollow structural sections are A500, rebar is A615 or A706, plate is A36, A572, A514 or A516. An "A36 square tube" is wrong on its face, and so is a mill certifying a product it does not roll.
3. Chemistry inside the claimed window. Every element on the row against the heat-analysis maximum for the grade. On a dual-certified heat, against every claimed grade. This is arithmetic on a published table and takes a minute.
4. Mechanicals make physical sense. Yield is below tensile on every real cert. Elongation carries its gauge length, 8 inches or 2 inches. If the order requires Charpy values, the row is present, with a test temperature. A992 has a yield ceiling as well as a floor, and a maximum yield-to-tensile ratio, because the range exists for seismic ductility.
5. The melt statement, if the job needs it. Federally funded highway work requires steel melted and manufactured in the United States, and the cert has to say so.
6. Dates and signatures in order. A mill report dated after the bill of lading that shipped the steel is a tell, not a clerical curiosity.

Ten tells a steel person catches on sight

These are the things that make an experienced buyer stop reading and pick up the phone. They are worth knowing because a system that is going to read certs for you has to catch every one of them, and most vendors' demos never mention them.

• Carbon printed to four decimals. Mills print two for most elements and three for phosphorus and sulfur. A value like 0.1624 was not printed by a mill.
• A chemistry row identical across two different heats to every decimal. Real heats vary in the third decimal of sulfur and phosphorus.
• Yield equal to or above tensile. Physically impossible.
• A chemistry value outside the claimed grade, shown as passing.
• Elongation with no gauge length.
• Charpy values with no test temperature.
• A heat number shaped like an invoice number, or one heat appearing on two different grades' certs without dual certification.
• A mill report dated after the bill of lading.
• A grade on the wrong product form.
• A section that does not exist. The wide-flange series is a discrete list; a W12x33 is not in it.

The limits my gate enforces, for the three grades that cover most structural work

These are the heat-analysis values the chemistry gate in my recorded steel demo checks against, verified against the specification text on August 4 and 5, 2026: A36 Table 2 and Table 3, A572 Table 2 and Table 3, and A992 Table 1 and Table 2, with the A992 carbon equivalent limit per the AISC engineering FAQ. They are a working reference, not a substitute for the edition you certify to. Where I could not verify a value against the specification, the gate refuses to enforce it and routes the cert to a person instead of inventing a violation. An unknown beats a disputed number, everywhere in this work.

Notes. A36 manganese and silicon requirements depend on product form and thickness, and the carbon maximum steps up for heavier plate, so the table shows only the forms the demo corpus uses. A572 Grade 50 requires a manganese minimum, a manganese-to-carbon ratio of at least two to one, and exactly one of the microalloy Type provisions, columbium or vanadium or both, so a Grade 50 cert with no columbium and no vanadium is not conforming. A992 requires silicon and the residual maximums for copper, nickel, chromium and molybdenum as well; nitrogen is a steelmaking-practice requirement and need not appear on the report. The IIW carbon equivalent is carbon, plus manganese over six, plus chromium, molybdenum and vanadium over five, plus nickel and copper over fifteen.

Two things this table makes visible. An A36 plate cert at carbon 0.26 is out of specification even though the shapes column passes it. And a heat certified A36 and A572 Grade 50 at carbon 0.25 is legal for A36 and illegal for Grade 50: the cert's own row refutes half its claim, with no judgment involved.

Where automated reading fails

Vendors selling cert automation publish accuracy figures without a denominator, and the figure is usually measured on clean PDFs straight from the mill. The cert that reaches a fabricator is rarely that document. It has been printed, faxed to a distributor, rescanned, attached to an email, printed again and photocopied for the job file. I have run readers of both kinds, a classic OCR engine and a vision model, down exactly that copy ladder, and the measured results are being published with a trade article this November. Until then, here is what the pattern is, without the numbers.

• Thin-stroke loss. A fax of a fax drops the thinnest strokes first. An 8 becomes a 3, a 0.023 becomes 0.028, a decimal point disappears. The page still looks clean at arm's length, which is the problem.
• Plausible misreads. A classic OCR engine produces garbage when it fails, and garbage is easy to catch. A model produces a plausible number when it fails, and a plausible wrong sulfur value passes every sanity check except the one that compares it to the grade window.
• Row and column shifts. Multi-heat certs list several heats in a grid. Shift one row and every value is real and belongs to a different heat.
• The transfer-cert retype. A distributor's retyped cert is a fresh set of human keystrokes on top of the mill's. Published human keying studies put cell error rates in the range of one to five percent, and people catch only about half to four fifths of an existing error when checking by eye (Panko, on human error rates).
• Confident and wrong is the dangerous class. A reader that abstains costs you a minute. A reader that returns a wrong value with no flag costs you a heat, a truck, or an audit. The visibly wrecked copy tends to trigger caution; the copy that only looks clean is where confident errors concentrate.

The defense is not a better reader. It is two independent reads that must agree, a deterministic comparison of every value against the grade window, and a person on every disagreement. A single reader reporting its own confidence is a system grading its own homework.

What a chemistry gate does, and what mine caught

A gate is the check between allocation and the truck. It makes no model call, because every condition is a lookup or a comparison and a compliance check must not be arguable: heats on the pick list are a subset of heats with certs on file; every chemistry value is inside every claimed grade's window; carbon equivalent recomputed by the IIW formula; mechanicals inside their ranges by test location; Charpy present when the order requires it; the melt statement present when the job requires it.

I built one, put a fleet of agents around it to work a steel order desk end to end, and recorded the runs. In the run that now ships as the default, 35 orders went through, the gate checked 20 certificates and made 327 comparisons, four certificates were blocked with the element named, twelve orders were routed to a person with a specific question, nineteen shipped and were verified at the dock, one short pay was caught in reconciliation, and the model spend for the whole run was $1.75. Six recorded runs are published, including the worse ones, with the misses named.

The four blocked certificates, each carrying exactly one planted defect, are the point of this page:

• A heat certified A36 and A572 Grade 50 with carbon 0.24. Legal for A36, over the 0.23 maximum for Grade 50. The cert refuted its own dual certification.
• An A992 heat with yield 67 ksi. The floor is 50, and a check written as "yield at least 50" passes it. The ceiling is 65 for a flange-tested shape, and the range exists for seismic ductility.
• An A36 transfer cert with sulfur 0.063 against a 0.05 maximum. Either mistyped or altered; both require the original mill cert before shipment.
• An order requiring Charpy values allocated to a heat whose cert has no impact row. Testing was never performed, and you cannot retest paperwork.

Every number in that paragraph was counted from the recording, not estimated. The mills, heats, customers and prices in the demo are invented, and the page says so first. Watch the recorded run, including the certificate that refuted itself.

A receiving checklist you can print

1. Heat on the tag equals heat on the cert equals heat on the bill of lading.
2. The specification and edition on the cert match the purchase order, and the grade fits the product form.
3. Every chemistry value is inside the heat-analysis window for every grade claimed.
4. Carbon equivalent recomputed from the row, not read from the cert, when weldability matters.
5. Yield below tensile; elongation with its gauge length; Charpy with a temperature when ordered.
6. Melted-and-manufactured statement present on Buy America work.
7. Printed precision looks like a mill's: two decimals, three for phosphorus and sulfur.
8. No two heats share an identical chemistry row.
9. Cert date on or before the ship date.
10. On a transfer cert, the original mill report is on file or on the way, and nothing certified ships before it arrives.

Questions buyers ask

What is the difference between a mill test report and a certificate of conformance?

A mill test report carries measured results: the chemistry and mechanical tests the mill ran on that heat. A certificate of conformance is a statement that material meets a specification, usually issued by a distributor and referencing the mill's report. The distributor may add a conformance statement. It may never alter the mill's test data, and a conformance certificate with no mill report behind it is a promise, not a record.

What does dual certification on a mill cert mean?

One heat certified to more than one specification, such as A992, A572 Grade 50 and A36 on a single wide-flange line. It is normal and useful. The rule is that the heat must satisfy every claimed window at once, chemistry and mechanicals. Each added grade can only shrink the legal region, so a heat can be legal for A36 and illegal for A572-50 on the same printed row.

What is carbon equivalent and why is it on the cert?

A single number that summarizes how the heat's chemistry affects weldability. The common formula is the IIW one: carbon plus manganese over six, plus chromium, molybdenum and vanadium over five, plus nickel and copper over fifteen. A992 carries its own limit in the specification, 0.45 for flanges up to two inches and 0.47 above. For other grades a CE limit is guidance unless your customer invokes one, so treat a printed CE as something to recompute, not something to trust.

Can AI or OCR read mill certs accurately?

On a clean print, yes, close to perfectly. On the copies a fabricator actually receives, faxed, rescanned, photocopied, the failure that matters is not the value the reader refuses. It is the value it returns confidently and wrong, such as a 0.023 read as 0.028 with nothing flagged. The only reliable defense is a second, independent read and a deterministic check of the values against the grade window, with a person on every disagreement. A single reader that reports its own confidence is not a check.

What should I do when a cert's own numbers fail the grade it claims?

Quarantine the heat and go back to the source. If it is a mill cert, the mill has to reconcile it. If it is a transfer cert on a distributor's letterhead, ask for the original mill report before anything ships, because the retype step is where both honest typos and dishonest edits enter. Do not ship on the basis that the steel is probably fine. The paper is the product on certified work.