The Test Is Only as Good as the Sample Behind It

Every number on a certificate of analysis describes a sample. The batch it is meant to describe is several orders of magnitude larger, sits in containers that were filled in sequence, and has been vibrated across an ocean. The procedure that connects the two is sampling, and it is the least examined step in the entire quality chain.

That neglect has a reason: sampling problems are invisible. A bad weighing produces a suspicious number. A biased sample produces a perfectly ordinary number that happens to describe the wrong thing.

Two errors, and the larger one is unmeasured

Total variance in a reported result is the sum of sampling variance and analytical variance. Analytical variance is studied, documented and controlled: methods are validated, instruments qualified, replicates run. Sampling variance is assumed away, on the grounds that the powder was blended and is therefore uniform.

The assumption is testable and rarely tested. For a blended powder, the sampling variance across containers and depths is commonly several times the analytical variance. That means a laboratory could halve its analytical uncertainty tomorrow and change almost nothing about the confidence a buyer can place in the certificate.

What a plan actually specifies

Element	The decision it encodes
Number of containers sampled	How much of the batch the sample set can speak for
Position within each container	Whether segregation is visible or hidden
Individual or composite	Whether the report describes the average or the spread
Tool and technique	Whether the sampling itself changes the material
Sample size and retention	Whether a dispute can be investigated later

None of these elements is difficult. All of them are decisions, and a plan that does not state them has left them to whoever is holding the scoop on the day.

The square root rule and its limits

The traditional plan takes samples from the square root of the number of containers, rounded up, with a minimum of one per container for small batches. It is a workable convention and it encodes an assumption worth naming: that the containers are interchangeable.

They are not, always. A batch discharged from a blender into containers in sequence has a beginning, a middle and an end, and the fractions differ slightly across the discharge. Segregation during filling puts fines toward one part of the container and coarse particles toward another. A plan that samples the top of every drum certifies the top layer of every drum.

The correction is structural rather than numerical: sample across the discharge sequence, and sample at depth with a thief rather than from the surface. Both cost minutes. The thief also has a drawback worth knowing: it disturbs the pack, so the containers it entered need to be identified, and the sampling record should say so.

Composite or individual

A composite sample is blended from the individual samples and tested once. It answers the batch-average question cheaply, and it is blind to the spread. Two batches can produce identical composite results while one of them varies by a factor of two between containers.

Where the customer's process is sensitive to the spread, the plan should test some samples individually, even if only on one parameter. The usual compromise is composite testing for assay and individual testing for the parameter that varies most, which for a milled powder is often particle size.

The decision is not about analytical philosophy. It is about which question the customer is actually asking, and the batch history is the place where that answer becomes visible over time.

What the record owes the reader

A sampling record that says "sampled per SOP" has recorded a habit. The useful record names the containers, the positions, the tool, the quantity, the sample code assigned and the disposition of the remainder. Five minutes of writing, and it is the only thing that allows a later dispute to distinguish a bad batch from a bad sample.

The disposition point is practical. Sampling disturbs the pack, and a container that was opened, thief-sampled and resealed is not identical to one that was not. If those containers are shipped without a note, the customer's own incoming check will find the opening and raise a question that a sentence in the record would have answered.

Sampling at the customer's end

The same logic runs in reverse. The customer's incoming sample describes the container it came from, and the container it came from is one of twenty. Where the customer's use is critical, the incoming plan mirrors the supplier's: more than one container, more than one depth, and a note of what was taken from where.

The two plans should agree in design, even if they differ in scale, because disagreements between a supplier's certificate and a customer's incoming test are far more often sampling disagreements than analytical ones. The first question in any such dispute is therefore not about the method. It is about which gram of the batch each laboratory measured, and the certificate itself usually does not say.

Where this leaves a supply-chain partner

We treat the sampling plan as part of the product: written, versioned and available to the customer on request. Where a batch is blended and discharged in sequence, the plan covers the discharge, not just the container count. Where sampling disturbed the pack, the record says which containers.

That is a slower answer than "sampled per SOP", and it is the difference between a certificate that describes a sample and one that can be defended as describing a batch.

Reading a plan against the physical batch

A plan is tested against three physical facts about the batch, and the first is the discharge sequence. A blender discharges its content over minutes, and the fraction that leaves first differs slightly from the fraction that leaves last, because fines and coarse particles do not move together through a discharge valve. A plan that samples containers one, five and ten of a twenty-container batch has covered the sequence. One that samples containers one through five has covered a fifth of it.

The second is the container geometry. A drum filled with powder and transported across an ocean has been vibrated for weeks, and vibration moves fine material downward and coarse material upward, which is the reverse of what filling does. The plan's depth instruction should acknowledge that the container the laboratory receives is not the container that left the site.

The third is the tool. A thief that reaches the bottom of a drum solves the depth problem and creates a traceability one, because the containers it entered have been opened. The record should identify them, and the customer's incoming check should expect to find the evidence.

What the standards actually say

The general expectations for sampling sit in the ICH Q7 framework, which requires written procedures and appropriate sampling, and the European interpretation is in the GMP text. The compendial perspective, including the statistical tradition behind container sampling, is described in the USP general chapters. None of the three prescribes the square root rule as mandatory; all three require that the plan be written, justified and followed.

That gap between requirement and prescription is where the discipline lives. A supplier can satisfy the requirement with a one-line procedure and a scoop. The difference from a real plan is invisible until a batch is questioned, and by then the sample is gone, the drum is empty, and the only remaining evidence is the record.

The two-question version

Everything above compresses into two questions that a buyer can ask without reading the procedure. How many containers, and from what depth, did the sample behind our last certificate come from? And does the plan sample across the discharge sequence or only across the container count?

A supplier who answers both without hesitation has a plan that somebody thought about. A supplier who answers "per SOP" has a document, and the difference between the two is the difference between a certificate that describes a batch and one that describes a habit.

When the two parties sample differently

The most common supplier-customer disagreement is not about the method. It is about two different grams of the same batch, taken from two different containers, at two different depths, six weeks apart. The numbers differ by a tenth of a per cent and both laboratories are correct about the material they measured.

Resolving it requires reconstructing both samples, which is why the sampling record has to be as specific as the analytical one. Which container, which depth, which tool, which date. With four facts per side, the disagreement usually resolves into an explanation: the customer sampled the container that had been at the back of the pallet, or the one opened for sampling at dispatch, or a container from a different discharge position.

Without those facts, the conversation defaults to method comparison, and method comparison is the wrong instrument for a sampling problem. Two laboratories can spend a fortnight comparing chromatograms that describe different powder.

The preventive measure is a sampling plan agreed between the parties before the first shipment, at least in outline, so that the incoming sample and the release sample are drawn on comparable principles. It costs one exchange and removes an entire category of dispute.

The record that makes it possible

Four fields turn a sampling record into evidence: container identifier, depth or position, tool and date. A fifth field, the disposition of the remainder, is what protects the next batch, because it names the containers that were opened. None of the five requires new equipment, and together they are the cheapest traceability improvement available to a powder operation.

One question tests whether the record exists in practice: which containers were sampled for our last certificate, and were any of them shipped? A supplier with the answer has the record. A supplier without it has a procedure that says the record exists.

Frequently asked questions

How many containers should be sampled in a batch?

The traditional rule is the square root of the number of containers, rounded up, plus one for small batches. It is a convention rather than a law, and it assumes the containers are interchangeable. Where a batch was blended and discharged to containers in sequence, sampling should cover the beginning, middle and end of the discharge regardless of the count.

Is a composite sample better than individual samples?

A composite answers the batch-average question and hides the spread. Individual samples answer the spread question at a higher testing cost. The plan should state which question it is answering, because the two answers lead to different decisions about the same batch.

What is the most common sampling error for powders?

Taking every sample from the accessible part of the container, usually the top. Powders segregate during filling and transport, so the top layer is not the batch. A sampling thief that reaches depth, or a plan that requires opening and cone-blending before sampling, addresses it.

Does sampling matter more or less than analytical precision?

More, usually. For a well-mixed powder the analytical method may have a relative standard deviation under one per cent while the sampling variance across a batch runs several times that. Most quality systems spend their effort on the smaller of the two errors.

What should the sampling record contain?

Which containers, which positions within each container, the tool used, the quantity taken, the sample's assigned code, and what happened to the remainder. The last point matters because a sampling procedure that disturbs the pack creates a deviation every time it is used.

Can a customer request the sampling plan?

Yes, and it is a reasonable request. The plan is a controlled document, it does not reveal commercial information, and it is the bridge between the batch and the certificate the customer relies on.