An interlaboratory comparison nobody had run

The headline finding is that within-laboratory repeatability was good in every case and between-laboratory reproducibility was not, and that the entire gap is explained by disclosed method choices rather than by anything hidden. Each service answered the question it was asked using the method it publishes. Each answer is defensible. The spread between them is nonetheless larger than the differences vendors compete on in their marketing, which means the number a buyer is comparing is not a property of the material.

The sample is not the batch

The most important sentence in every third-party report in this market is the one identifying what was tested, and it always says the same thing: the sample submitted. That phrasing is precise and correct, and the entire trade reads past it. A result obtained on one vial extends to a batch only if the vial is representative, and representativeness is a property of how the vial was selected, not of how carefully it was analysed.

In regulated practice, sampling is a controlled activity in its own right: the accreditation standard treats it as part of the laboratory activity, requiring a documented sampling plan and records of how the portion tested was obtained.¹ Where a laboratory receives a sample it did not draw, the standard expects the report to make clear that the results apply to the sample as received. All four services do this. The market quotes them anyway as though the batch had been tested, which inverts the pharmacopoeial convention that a result on a sample is evidence about a batch only under a stated sampling assumption.²

The practical significance depends on the fill. A batch filled in a single session from a homogeneous bulk solution is likely to be uniform, and a single vial is decent evidence about it. A batch assembled from subdivided bulk, filled across sessions, or blended from more than one synthesis is not, and a single vial is evidence about a vial. Nothing on a report tells a reader which situation applies, because the laboratory does not know either.

A worked estimate of the selection effect

Since the unpublished results are invisible, any estimate of the selection effect has to be constructed rather than measured, and the Journal offers the following as an illustration rather than a finding. Suppose the true distribution of purity results for a competent supplier is centred at 98.0% with a standard deviation of 0.8 points, which is consistent with the spread we observe on repeated submissions. Suppose the supplier publishes results above 98.0% and files the rest.

The published mean is then approximately 98.6%, the published minimum is 98.0%, and the apparent variability is roughly halved. A buyer reading the published set would conclude that the supplier’s process is both better and more consistent than it is, and would be wrong on both counts without anybody having lied. Increase the publication threshold to 98.5% and the published mean rises to 99.0% while the true mean is unchanged.

The arithmetic is elementary and the point of doing it is to show how modest an amount of selection is needed to produce a large apparent effect. No fabrication, no dishonest analyst, no altered document: one decision about which reports to circulate. Any market whose evidence base is assembled from voluntarily disclosed tests commissioned by interested parties has this property, and the remedy is structural rather than moral.

When a vendor commissions testing of its own product

The clearest case of the conflict is also the most common. A vendor wishing to advertise independent verification commissions a test, selects the sample, pays the invoice, receives the report and publishes it. Each of those steps is ordinary commerce and none of them is improper. Taken together they mean the verification of a product has been arranged and funded by the party selling it, which is a conflict of interest in the plain sense of the term.

It is worth being exact about what the conflict does and does not imply. It does not imply that the laboratory’s measurement is wrong; there is no reason to think it is, and the laboratory has strong reputational reasons for accuracy. It does not imply that the vendor is dishonest. What it implies is that the result is not independent of the vendor’s interests in the way the word independent suggests, because the vendor controlled the sample, the timing and the disclosure.

The comparable arrangement in other industries — a manufacturer paying an accredited laboratory to test its own product — is normal and is managed by requiring that results be reported to a regulator or a purchaser with enforcement power, so that disclosure is not optional. Neither exists here. The Journal therefore labels vendor-commissioned results as such wherever it cites them, which is the only lever a publication has.

The blind duplicate exercise

Twelve vials, one lot, purchased at retail without disclosure of purpose. Two vials were sent to each of the three assay services under two different submitter names and addresses, so that each laboratory received two nominally unrelated submissions of the same material some three weeks apart. Six further vials were retained. Each service was asked for its standard purity determination at its standard price and turnaround, with no special instructions.

The design tests two distinct quantities that the trade conflates. Repeatability is the agreement between duplicate determinations within one laboratory; reproducibility is the agreement between laboratories. Interlaboratory studies in analytical chemistry consistently find the second to be substantially worse than the first, and the variance decomposition that separates them is standard methodology.³ The distinction between repeatability and intermediate precision is formalised in the validation guidance,⁴ and multi-site studies in adjacent fields have repeatedly found between-laboratory agreement on identical samples to be the harder problem.⁵

All three services were informed after the fact, before publication, and each was given the opportunity to comment on its own method as printed and on the comparison as a whole. All three responded. Two supplied additional method detail that has been incorporated. One disputed the framing of the comparison, and its objection is printed in the correspondence below. None of the three asked for its result to be withheld, which the Journal records because it did not have to be that way.

What came back

Within-laboratory repeatability was good. The two determinations from each service agreed to within 0.3 percentage points in every case, and to within 0.1 in one, which is about what a well-controlled chromatographic method should deliver on duplicate material and is a genuinely reassuring result.

Between-laboratory reproducibility was another matter. The three services returned figures spanning 2.1 percentage points on material from one lot. Every point of that spread is accounted for by disclosed method differences: gradient duration, integration threshold, the retention-time cut-off defining the solvent front, and whether an orthogonal second gradient was run and the lower figure reported. Rerun the raw data from the shallowest method with the fastest method’s integration threshold and the two figures converge to within 0.4 points, which is the strongest available demonstration that the disagreement is methodological rather than analytical.

Identity results agreed completely: all three found a single dominant species at the expected mass, and none reported evidence of an unrelated compound, which is the ordinary outcome of intact-mass confirmation on submitted material.⁶ The two services reporting peptide content returned 93% and 91% of label, a difference within the stated uncertainty of nitrogen determination. The material, in short, was what it claimed to be, and the disagreement was confined to the second significant figure of the number the market competes on.⁷

What this exercise cannot establish

Four limitations, stated because the alternative is letting readers over-read a small study. First, one lot of one compound from one supplier is not a sample from which the performance of these services in general can be inferred; it is an existence proof about method-driven spread. Second, three services is too few for any statistical treatment beyond the descriptive; published round-robin studies of peptide purity use seven or more participants for exactly that reason.⁸

Third, and most important, the exercise tested reproducibility, not accuracy. All three could be equally wrong: without a certified reference standard of known purity, there is no true value against which to score them, and the compendial approach to validating a purity procedure requires exactly such a reference to establish accuracy rather than mere agreement.⁹ What we measured is dispersion around an unknown centre, and the same constraint applies to any quantitation attempted without a matched standard.¹⁰

Fourth, blind submission tests a laboratory’s ordinary process, which is the point, but it also means we bought the cheapest standard product from each service rather than the most thorough. A comparison of each service’s best available package would be a different and probably more flattering study, and it would tell a buyer less, because almost nobody buys the best available package.

The Journal will repeat the exercise annually with a different compound and, funding permitting, against a certified reference standard. The design is published so that others can run it.

The archive, and why it matters more than the method

A public, searchable archive of prior results changes the epistemics of this market more than any improvement in analytical method would. The reason is the selection problem. A single report is a snapshot chosen by whoever paid for it. A record of every submission relating to a vendor, published whether or not the result flattered the submitter, is a series — and a series cannot be curated by choosing which snapshots to show.

Two of the four services maintain something of this kind. The Journal would like all four to, and has asked. The principal objection raised in response is a real one: client confidentiality is an obligation, not a preference, and a laboratory cannot unilaterally publish a paying client’s results. We accept that entirely, which is why our request is narrower. Publish the fact of the submission — the date, the vendor named by the submitter, and the compound — while keeping the result confidential where the client requires it.

That narrower version defeats most of the selection effect at no cost to confidentiality. A vendor with nine public submissions and three published results is visibly a vendor with six results it declined to publish, and a reader can price that in. It is, so far as we can tell, the single highest-leverage change available to this sector, and it requires no new equipment and no new method.

What a badge asserts, and what it supports

A verification badge on a product listing typically asserts, in one or two words and a graphic, that the product has been tested by a named service. Consider what that claim leaves open. Which lot was tested. When. Who submitted the sample and how it was selected. What was measured — purity alone, or identity, or content. By what method. Whether the lot currently on sale is the lot that was tested. Whether the report is available to the reader.

Every one of those is material and every one is absent. A badge is therefore not evidence; it is an assertion that evidence exists, offered without the particulars that would let anybody assess it. The Journal does not treat badges as evidence in its coverage, states so wherever it reports one, and will not cite a badge as support for a claim about material.

The remedy is four data points the vendor already possesses: the lot number, the date of analysis, the service, and a link to the report. Several listings in this market already carry them, which establishes both that it is possible and that it is not commercially fatal. What it requires is that a badge become perishable — attached to a lot rather than to a product line — and perishability is precisely the property a marketing asset is designed not to have.

Our own position, stated in full

Janoshik Analytical and PeptideMeter both advertise in The Compound Journal. Both relationships are disclosed by name on our funding page, together with every other sponsor. No advertiser sees editorial copy before publication, no advertiser has any role in commissioning or reviewing coverage, and the analytical-chemistry desk is contractually barred from consulting for any vendor, testing service or compounding pharmacy. This article was edited by the standards desk under the same rules as every other piece in the department.

The Journal also pays these services. We have submitted samples to three of the four on commercial terms, at list prices, and the blind duplicate exercise described above was funded from editorial budget. We are therefore simultaneously a customer of the institutions we are reporting on and a recipient of advertising revenue from two of them. Readers are entitled to weigh that, and the only useful response we can offer is to state it plainly and to publish objections.

Our position on the substance is unchanged by any of it. All four services are legitimate operations and we have no evidence of dishonesty by any of them. The problems this article describes are structural — who commissions testing, who decides what is published, and what a sample can support about a batch — and they would persist unchanged if every person working at all four organisations were beyond reproach. Correspondence to standards@compoundjournal.com.

What the Journal thinks these services are worth

A summary judgement, since a critical article of this length invites the inference that we think the sector is worthless. We do not. Independent testing in this market is the only mechanism by which a buyer can obtain information about material that is not supplied by the party selling it, and its existence is the difference between a market with some evidence in it and a market with none. Several of the reports these services produce are better documents than the manufacturer certificates they are checking, which is a low bar cleared with room to spare.

The three criticisms we would press are narrow. A sample is not a batch, and the trade cites samples as batches. The party paying for a test decides whether anybody sees it, and the visible corpus is therefore selected. And a badge on a listing has dropped every particular a reader would need. None of these is an analytical failing and none is a failing of the services in isolation; the second and third are properties of the market that surrounds them.

What we would tell a reader is this. A third-party report on a vial you selected and posted yourself is strong evidence about that vial. A third-party report published by the vendor is weaker evidence, of an amount you cannot determine. A badge is not evidence. And nothing in any of the three is a statement about whether anybody should administer the contents to anything.

The Journal’s summary of this sector is that it is better than the market deserves and weaker than the market believes. Three laboratories and an auditor, selling individual tests mostly by post, mostly to private individuals, produce documents that are frequently better than the manufacturer certificates they check. What they cannot produce, because no institution in this market can, is a result that the party being examined does not control the disclosure of.