Acceptance Criteria: Whether They Came From Data or Habit

Ask where a limit came from and the answers fall into four groups. It is in the monograph. It was set from our own batch data. The customer's process requires it. Or: it has always been that number. The fourth answer is the most common and the least examined.

A limit is a decision about risk, expressed as a number. Whether that decision was made deliberately is a separate question from whether the number is met, and it is the question that determines whether a specification means anything.

Four legitimate origins

Origin	What it asserts	Its weakness
Compendial monograph	A public standard exists for this substance	Silent on impurities the monograph does not name
Statistical study of the process	The process reliably achieves this level	Describes the process at the time of the study
Downstream requirement	The customer's process needs this	Requires the customer to have characterized their process
Regulatory or safety threshold	An authority has decided	Not available for most impurities

The second origin is the one that most specifications claim and few can document. A statistical limit requires a distribution: a mean, a standard deviation, and a chosen distance from the mean that the process achieves with an acceptable failure rate. That is a calculation, and calculations leave traces.

What a data-based limit looks like

Take twenty consecutive batches of a related substance whose results average 0.04% with a standard deviation of 0.012%. A limit at the mean plus five standard deviations is 0.10%. That limit will be met by roughly one batch in a million if the process stays where it is, and it will be breached immediately if the process shifts, which is exactly what a limit should do.

Now take the same twenty batches and read the highest result: 0.08%. Set the limit at 0.10% and it looks the same, but it is not. The second limit was set from a maximum, not from a distribution, and a maximum carries no information about what happens next.

This is why the difference between "our highest result plus a margin" and "our mean plus a multiple of the standard deviation" matters. The first is a description of the past. The second is a prediction, and it can be tested as batches accumulate.

The copying problem

Specifications travel. A limit set for one product is reused for a related one, then for a different grade, then for a different market. Each transfer is invisible in the document because the number looks the same.

The consequence is a specification that is neither protective nor achievable, and nobody can say which. A limit of 0.10% on an impurity that the process produces at 0.02% is not protective, it is decorative. A limit of 0.10% on an impurity the process produces at 0.09% is a permanent crisis.

The gap between a supplier's specification and a customer's in-house specification is where this becomes visible. Two documents, the same molecule, different numbers, and neither party able to explain the difference because neither number was derived.

The asymmetry of getting it wrong

A limit set too tight produces a steady stream of results that pass but sit close enough to the edge to require justification. Investigations proliferate, each one concluding that the batch is fine. The real cost is not the investigations. It is that the organisation learns to treat a limit as an obstacle rather than a signal.

A limit set too loose produces nothing at all until the customer's process fails, at which point the variability that the specification accepted becomes the customer's problem. The supplier's certificate is compliant throughout, which is the worst position for the customer, because there is no defect to point at.

Both errors are made by default rather than by decision, and both are cheap to avoid at the point of setting the specification.

Setting a criterion that can be defended

Four steps, in order. Establish what the customer's process actually needs, in numbers, from the customer. Characterize the process's own distribution from a stable period, with enough batches to estimate a standard deviation. Choose the distance from the mean that reflects the consequence of failure, not a round number. And confirm the method can measure at that level with an uncertainty small enough to make the decision meaningful.

The last step is the one most often skipped. A limit of 0.05% on an impurity whose quantitation limit is 0.03% and whose uncertainty at that level is plus or minus 0.02% is a limit that cannot be applied. The arithmetic of a result sitting at a limit is where this becomes concrete, and it is also where the conversation with a customer becomes productive, because the number stops being a preference and becomes an interval.

Changing a limit later

Limits move in one direction easily and the other direction with difficulty. Tightening is a change the customer welcomes and the supplier pays for. Widening is a change that requires justification, and the justification has to predate the batch that prompted it.

The test an auditor applies is simple. Was there data supporting the change before the non-conformance, or was the change made to accommodate it? A specification revised after a failure, with no earlier evidence, is treated as a dilution of the non-conformity regardless of how the paperwork describes it. The residue test is a useful example, because its limit is often set by habit at a value the process has never approached.

What this means in a supply conversation

We state, for each limit we apply, whether it is compendial, derived from our own batch distribution, or set by the customer's requirement, and we say which. Where a customer's limit is tighter than ours, we treat that as their decision about their process and we say whether we can meet it with margin.

The framework for specification setting, including the distinction between tests that are universal and tests that are product specific, is set out in ICH Q6A, with the compendial texts maintained by EDQM and the USP.

The question to ask a supplier is not what the limit is. It is: what data set was this limit derived from, and what would you do if a batch came in at ninety percent of it?

A worked example, with illustrative numbers

Twenty consecutive batches of a related substance average 0.042% with a standard deviation of 0.011%. A limit at the mean plus five standard deviations is 0.097%, which is usually rounded to 0.10%. The rounding is harmless. The basis is not: the limit is now a statement about the distribution, and the distribution can be re-tested as new batches arrive.

The same twenty batches have a maximum of 0.068%. A limit set at the maximum plus a margin lands near 0.08%, and it looks tighter and therefore safer. It is neither. It is a limit that has no arithmetic behind it, so it will either be breached by a normal batch or it will hold for reasons nobody can state.

Capability is the number that makes the difference visible. The distance from the mean to the limit divided by three standard deviations gives an index; a value near one means the limit sits where the process occasionally reaches, and a value near two means the process has room. A specification without a capability figure is a limit with no evidence about whether it can be met.

When the customer's requirement is the binding one

Frequently the customer's process sets a tighter limit than the supplier's distribution would justify, and that is legitimate: their process may be less tolerant than the specification assumes. The consequence is that the supplier is now manufacturing to a capability requirement rather than to a limit, and the two parties should say so.

This is where the gap between the two documents becomes operational rather than administrative. If the customer's limit is 0.06% and the process runs at a mean of 0.042% with a standard deviation of 0.011%, the customer's limit is roughly 1.6 standard deviations away, and a normal batch will breach it occasionally. That is a conversation about capability, and it should happen before the first purchase order rather than after the third complaint.

What belongs in the specification document

For each limit: the origin, the data set it was derived from, the date of that derivation, and the capability at the time. Four lines per test. A specification with those four lines can be reviewed, challenged and revised rationally. A specification with a number and a unit cannot be reviewed at all, only obeyed.

One further test is worth writing into the specification: the review date. A limit derived from data collected before a process change describes a process that no longer exists, and a specification without a review date never notices. An annual look at whether the distribution still matches the limit costs an hour and prevents a limit from becoming folklore.

Frequently asked questions

Where do acceptance criteria legitimately come from?

Four sources. A compendial monograph, where the product is described by one. A statistical study of the process's own distribution, where the limit is set at a distance from the mean that the process achieves reliably. A downstream requirement, where the customer's process cannot tolerate more than a certain level. And a safety or regulatory threshold, where an authority has set the number.

What is wrong with copying a limit from a previous product?

Nothing, if the two products have the same impurity profile, the same process capability and the same downstream use. Usually they do not, so the copied limit is either tighter than necessary, which generates false failures, or looser than necessary, which provides no protection. Either way, nobody knows which.

How many batches are needed to set a statistical limit?

Enough to estimate the distribution, which in practice means a minimum of ten to twenty and preferably more, from a process in a stable state. A limit derived from three batches describes three batches. A limit derived from batches made during a campaign that has since changed describes a process that no longer exists.

Can a limit be widened later?

It can, through change control, with justification. Widening a limit in response to a failing batch is the pattern regulators treat as dilution of a non-conformity rather than a specification change, and the distinction is whether the change was justified by data collected before the failure.

What is the cost of a limit that is too tight?

Recurring false failures, each of which consumes an investigation, and a slow drift toward testing to the limit rather than to the process. A limit that the process meets only by sorting is a limit that has been set in the wrong place.

What is the cost of a limit that is too loose?

The customer absorbs the variability. A specification wide enough to accept everything provides no signal, and the downstream process discovers the variation instead, at a point where it is more expensive and harder to attribute.