A buyer qualifying a new powder supplier will nearly always ask for three consecutive production lots. We went looking for the standard it comes from and could not find one. What we found was a footnote.
If three lots is a convention rather than an obligation, there is no leverage in demanding the number — most suppliers will agree to it. The leverage is in specifying what those lots have to show: how material comes out of the drum, how it is split down to the few grams that reach an instrument, which method a number belongs to, whose laboratory wins a disagreement, and what happens to what is left over. All of it has to be written down before the first sample ships.
All of it, because a qualification sample is not a claim about the powder in the jar. It is a claim about the process that made it, and a jar can be filled in ways a production lot cannot. The automotive answer is procedural: the AIAG production part approval process requires parts from a significant production run — reported to us second-hand as at least 300 consecutive parts, at the production site and rate, using the production tooling, materials and operators. Every clause exists to defeat a sample made specially. PPAP is automotive and binds no powder supplier, but its logic transplants: make the sample prove it came from production, through a lot number traceable to a batch record and a certificate of analysis (COA) carrying the fields a commercial lot would carry, by the same methods. That is why our 3Y-TZP powder page treats a grade as qualified only against COAs from three consecutive lots read to one specification.
The number came from a footnote
The 1987 Guideline on General Principles of Process Validation, prepared by FDA's Center for Drugs and Biologics and Center for Devices and Radiological Health and announced in the Federal Register of 11 May 1987 (52 FR 17638), is where the three-batch convention is usually said to begin. The word "three" does not appear in it. Not once. Its performance-qualification section says only that challenges "should be repeated enough times to assure that the results are meaningful and consistent."
A number does appear, in footnote (7) on page 10, as an example: "For example, the AAMI Guideline for Industrial Ethylene Oxide Sterilization of Medical Devices approved 2 December 1981, states: 'The performance qualification should include a minimum of 3 successful, planned qualification runs, in which all of the acceptance criteria are met ….. (5.3.1.2.).'" If the convention has a written origin, this is the closest thing to it that the record offers: a parenthetical illustration, quoting a third party's standard, about sterilising medical devices with ethylene oxide. It did not enter as a requirement, and it was not about materials.
That guideline is no longer FDA's position. The January 2011 Guidance for Industry — Process Validation: General Principles and Practices, from CDER, CBER and CVM, "replaces the 1987 guidance" and specifies no number of batches. It asks instead that the number of samples be "adequate to provide sufficient statistical confidence of quality both within a batch and between batches", and for "statistical metrics defining both intra-batch and inter-batch variability." This is pharmaceutical process validation, and FDA says its guidance does "not establish legally enforceable responsibilities."
The one regulatory text that names the number keeps it and demotes it. EU GMP Annex 15 (Qualification and Validation, Brussels, 30 March 2015, in operation from 1 October 2015) says at §5.20 that "without prejudice to 5.19, it is generally considered acceptable that a minimum of three consecutive batches manufactured under routine conditions could constitute a validation of the process", adding that an alternative number may be justified. Section 5.19, which §5.20 defers to, is the operative one: each manufacturer "must determine and justify the number of batches necessary to demonstrate a high level of assurance that the process is capable of consistently delivering quality product." Annex 15 is pharmaceutical GMP and does not reach a ceramics powder supplier — but even where the phrase is written into regulation, the duty to justify the number sits above it.
Elsewhere the number is absent: aerospace first-article inspection under AS9102 can be satisfied by a single representative part. We could not obtain IATF 16949 or the AIAG manuals, so we found no three-lot requirement rather than establishing that none exists.
One "three" is genuinely in a standard, and it is not the one anybody quotes. ISO 3954:2007 §3.1 requires that when sampling during continuous discharge, "at least three increments shall be taken" — one in the first third of the discharge, one halfway, one in the last third. Three increments from one lot: a rule about representativeness, not reproducibility. So the sentence our supplier due-diligence checklist could only assert now has its proof: nothing above buyer level requires three consecutive lots, and they happen only if the contract says so.
Three lots cannot tell you how much a process varies
The second reason not to treat the number as the point is arithmetic rather than provenance.
Say the three lots come back and you compute a standard deviation across them. For n lots, a 95 % confidence interval on the true lot-to-lot standard deviation σ runs from s·√((n−1)/χ²₀.₉₇₅,ₙ₋₁) to s·√((n−1)/χ²₀.₀₂₅,ₙ₋₁). At n = 3 there are two degrees of freedom, and a chi-square distribution with two degrees of freedom is exactly an exponential distribution with mean 2 — so the quantiles are closed-form and need no table:
- χ²₀.₉₇₅,₂ = −2·ln(0.025) = 7.378, giving √(2 / 7.378) = 0.52
- χ²₀.₀₂₅,₂ = −2·ln(0.975) = 0.05064, giving √(2 / 0.05064) = 6.28
Three lots put the true lot-to-lot standard deviation between roughly 0.52× and 6.28× what you observed — a twelve-fold span. A supplier whose real spread is six times what your three lots showed is entirely consistent with your three lots. Ten lots would narrow it to roughly 0.69× to 1.83×, and nobody buys ten qualification lots.
That is not an argument for demanding more. It is an argument about what three are for. They can show that a mean moved, that a lot failed a limit, and that the supplier can make the material more than once. They cannot characterise variability, so a protocol that computes a capability index from three points and reports it as a property of the supplier is reading precision into data that does not carry it.
What reaches the instrument is several reductions from the lot
The number on the report was never measured on the lot. It was measured on a few grams descended from it, through a chain that ISO 3954:2007 — a powder-metallurgy sampling standard, which does not govern ceramic powders — names step by step: lot → increment, what a sampling device takes at one time → gross sample, all the increments from one lot → composite sample, that blended → test sample → test portion, what the test is actually carried out on. Four or five reductions between drum and cuvette, none of them visible in the result, and each can change the answer, because the material does not stay mixed. ISO 3954 §4.1 notes that "demixing may occur at any time when a batch of powder is set in motion, for example when filling containers, emptying containers, during transportation or if subjected to vibration during storage." Under vibration coarse particles migrate upward and fines downward, which is why a scoop off the top of a drum that has crossed an ocean is not a sample of the drum. A zirconia or silicon carbide sample travels the same chain, and three of the standard's remedies transfer almost verbatim into a protocol:
- Sample while the powder is moving. §4.2.1 says that "whenever possible" sampling from a continuous discharge stream "should be preferred over" thief sampling from packaged containers, and §4.2.2 that the container be exposed to the stream and withdrawn so that all portions of the stream have an equal chance of entering it.
- If a drum must be sampled, the thief goes in at 0.7 of the radius. §4.2.3.4: for a single increment from a cylindrical container filled through an opening above its central axis, the thief "shall be inserted at a distance from the centre equal to 0,7 of the radius". Not the middle, not the wall — and it must reach every level.
- Cone-and-quartering is not on the list. §4.3 names four acceptable splitting devices: spinning riffler, sample splitter, rotary sample splitter, rotating-cone splitter. Cone-and-quartering, still the laboratory default in many places, is absent — and in the study both instrument makers cite for it — a sand-and-sugar mixture reported by Allen — cone-and-quartering came out around 6.8 against roughly 0.1 for a spinning riffler, a factor of about fifty, though the two sources disagree on the exact riffler figure and one notes the mixture is close to a worst case.
One standard prices the whole business with its own design. ISO 13320:2020 (Particle size analysis — Laser diffraction methods, second edition) defines two separate repeatability tests: instrument repeatability re-measures the same aliquot, method repeatability measures different sub-samples. Of the second, §3.1.22 says "the variability includes the variabilities of sub sampling technique, of the sampled material together and of the instrument." The gap between the two tests is the price of sub-sampling. An instrument maker's published reading of that edition — Microtrac's — allows the D50 within ±1.5 % on the instrument test and ±2.5 % on the method test; we could not read the clause itself, so we quote that pair only as Microtrac's reading. The same logic applies to your own laboratory: a buyer whose incoming inspection eats a third of the tolerance cannot resolve a shift smaller than its own noise. Which instruments to run on a received lot is a separate question, covered in our SEM/EDS lot verification method.
Every part of the protocol is contractual, or it does not exist
Nothing above obliges anyone. We looked for a non-pharmaceutical standard or Korean statute requiring retained physical samples of industrial materials and found none. For industrial materials, retained samples exist only because a contract says they do.
The most developed model is again pharmaceutical and again binds no powder supplier: EU GMP Annex 19, Reference and Retention Samples (Brussels, 14 December 2005). It settles four things a powder contract leaves open. It separates the reference sample, kept "for the purpose of being analysed", from the retention sample, kept "for identification purposes". It sizes the first by function — §4.1 requires enough to permit "on, at least, two occasions" the full analytical controls on the batch. It sets a period, §3.2 holding starting-material samples for at least two years after release of product. And §4.4 requires the analytical materials and equipment to stay available for as long as the sample is held. That is the clause buyers forget: a sealed jar you can no longer test is not a retained sample.
The retained sample also turns out to be the same clause as the one that resolves disputes. In base-metal concentrate, semi-refined metal and iron ore contracts an umpire assay runs like this: each side assays its own portion of the same lot's split sample and the results are exchanged; if they differ by more than a contractually agreed splitting limit, a third sealed portion — set aside at sampling time for exactly this purpose — goes to a pre-nominated umpire laboratory whose result binds, the cost falling on the party whose figure was further from it. It is contract custom rather than a published rule, so you draft it yourself. The structural point survives the drafting: an umpire clause is impossible unless a third sealed portion was set aside when the sample was taken. Retained samples and dispute resolution are one clause seen from two ends, and a buyer who writes the second without the first has written something that cannot run.
Nominating the laboratory settles whose number wins — in Korea, one accredited to KS Q ISO/IEC 17025 under KOLAS, whose reports carry through the ILAC and APAC mutual recognition arrangements. Naming the method settles which number is compared: a D50 by laser diffraction and a D50 by sieve are different numbers for the same powder, and we have no sourced figure for how far apart. Hence a certificate stating the method and dispersion conditions rather than the median alone, as ours do for green SiC micro powder.
What to fix before the first sample ships
- Sampling method and point — a moving stream where possible; otherwise how many containers, how many increments, and the thief's insertion position.
- The splitting device, named — a riffler or splitter, not cone-and-quartering.
- The test method for every specified property, with preparation and dispersion conditions.
- Retained sample — how much, sealed by whom, stored where and how long, with method and equipment kept available throughout.
- Nominated laboratory and umpire clause, with the splitting limit and who pays.
- Traceability to a production batch record, with production date and batch size.
- What each lot has to demonstrate — stated as questions, not as a count.
Grade, particle-size designation, impurity basis and packing belong before any of this; those are the pre-RFQ decisions in our SiC powder checkpoints, and the block itself can be assembled with our RFQ specification builder, which carries D50 with tolerance and method, trace metals on the element basis, and production-lot COAs as line items. One customs note for Korean buyers: under 관세법 제94조제3호 and 관세법 시행규칙 제45조제1항제3호 a sample may be exempted from duty as a 견본품 where its customs value is USD 250 or less. The wording is "면제할 수 있다": permissive, applied at entry.
Frequently Asked Questions
Is three consecutive lots an FDA or ISO requirement?
No standard we could locate requires it. The 1987 FDA guideline does not contain the word "three"; the number appears only in a footnote quoting a 1981 AAMI ethylene-oxide sterilization guideline as an example, and the 2011 guidance that replaced it specifies no number. EU GMP Annex 15 §5.20 does keep "a minimum of three consecutive batches", but as something "generally considered acceptable" without prejudice to §5.19's duty to justify the number — and it is pharmaceutical GMP, which does not bind a powder supplier.
If three lots is arbitrary, how many should we ask for?
That is the wrong quantity to optimise. Three lots cannot usefully bound lot-to-lot variability — the 95 % confidence interval on the true standard deviation runs from about 0.52× to 6.28× the value you observe. Ask instead what each lot has to demonstrate, and spend the effort on how the samples are taken, split, measured and retained.
Can we just take a scoop from the top of the drum?
No. ISO 3954:2007 notes that demixing occurs whenever powder is set in motion, including transport and storage vibration, with coarse particles migrating up and fines down. Sample from a moving discharge stream where you can; otherwise the thief must reach every level, and a single increment from a cylindrical container goes in at 0.7 of the radius from the centre.
Who pays for qualification samples, and how long do they take?
We could not source any credible norm for either — no typical cost split, no typical lead time — so we state none. Both are terms to settle in the same contract that fixes the sampling and retention rules; treat a supplier's answer as a quotation, not as the industry's.
References (Public Sources)
- US Food and Drug Administration, Guideline on General Principles of Process Validation, May 1987, prepared by the Center for Drugs and Biologics and the Center for Devices and Radiological Health; availability announced in the Federal Register of 11 May 1987 (52 FR 17638). The copy we read was not obtained from FDA's own site, the agency having withdrawn the document.
- US Food and Drug Administration, Guidance for Industry — Process Validation: General Principles and Practices, January 2011, CGMP Revision 1, issued by CDER, CBER and CVM.
- European Commission, EudraLex Volume 4, EU Guidelines for Good Manufacturing Practice: Annex 15 (Qualification and Validation), Brussels, 30 March 2015, in operation 1 October 2015, §§5.19 and 5.20; and Annex 19 (Reference and Retention Samples), Brussels, 14 December 2005, §§2.1, 3.2, 4.1 and 4.4.
- ISO 3954:2007, Powders for powder metallurgical purposes — Sampling, second edition, ISO/TC 119/SC 2 — the clause 2 vocabulary and the sampling and splitting provisions of clauses 3 and 4, read from a publicly available preview.
- ISO 13320:2020, Particle size analysis — Laser diffraction methods, second edition, ISO/TC 24/SC 4 — the definition of method repeatability at §3.1.22. The percentile tolerances are Microtrac's published reading of clause 6, not text we read.
- AIAG production part approval process material, reached only through a customer company's own PPAP procedure and quality-profession discussion. The manual is a paid publication we could not obtain, so the significant-production-run wording and the 300-part figure are second-hand throughout.
- SAE International / IAQG AS9102 first article inspection requirement, from secondary summaries rather than text we read; the revision letter and issue date are not verified.
- HORIBA and Malvern Panalytical, both publishing a white paper titled Sampling for Particle Size Analysis — the sample-division error comparison, attributed by each to a different edition of T. Allen, Particle Size Measurement, with figures that disagree. We did not read Allen's book and reproduce no values.
- Contract-clause banks, mining-software documentation, a copper concentrate purchase contract filed with the SEC, and a commercial laboratory's umpire testing service description — the umpire assay mechanism. We found no published model form.
- 국가법령정보센터: 관세법 제94조 and 관세법 시행규칙 제45조. Source renderings of the responsible ministry and of the enactment metadata conflict, so we cite article and item numbers only.
Positions and figures reflect public sources at the time of writing. Where an edition year is not given above, it is because we could not verify one.
Nami Tech Solutions works project by project on ceramic and rare-earth powders rather than from standing inventory, and the sample stage is where most of our work sits. We agree the sampling and splitting method with the plant before material is drawn, hold the specification's measurement methods to the same wording on the sample and on the commercial lot, and settle who retains which sealed portions, so that a later disagreement is resolvable rather than rhetorical.
