What FDA Inspectors Want From Your OOS File

By Aldo Xhango, Founder, GuideGxP

Read enough FDA warning letters and a pattern emerges that has little to do with laboratory technique. Investigators rarely cite a retest that was performed badly. They cite a retest that nobody justified in writing.

The regulation is short. 21 CFR 211.192 requires a thorough investigation of any unexplained discrepancy or any failure of a batch to meet its specifications, whether or not the batch has been distributed, and requires the conclusions and follow-up to be written down.¹ FDA's guidance on investigating out-of-specification (OOS) test results sets out how the agency evaluates that investigation.² Neither document asks for elegance. Both ask you to show your reasoning.

What follows is the reasoning an inspector reconstructs from your file, phase by phase, and the three places 2026 enforcement shows it breaking down.

The Pattern In The Citations

In an April 2026 warning letter to a U.S. manufacturer, the firm obtained a failing microbial count, could not identify the fungal species through its contract laboratory, concluded without evidence that contamination had occurred during handling, transportation, or testing, and then invalidated the failing result on the strength of resampling and released the lot.³ The same letter describes two stability results for subpotent assay invalidated on resampling data with no laboratory error identified and no rationale recorded.

A June 2026 letter to a manufacturer in South Korea documents something more deliberate. Investigating an unknown impurity, the firm prepared a second sample using glassware marked for disposal so that it would mimic the original failed sample. That sample passed, and the firm used the passing hypothesis test as its justification to invalidate the original result.⁴

A May 2026 letter to a European sterile manufacturer takes the third route: the investigations existed, but they lacked adequate scientific rationale to support the root causes they identified and the associated data were not documented contemporaneously.⁵

Warning letters are an imperfect mirror. They show what went wrong at other sites, not what is working at yours, and they arrive with the outcome already known. They are also the only public record of what an investigator concluded after reading a file cold, without the benefit of anyone in the room explaining it, which makes them the closest thing the industry has to a marking scheme.

Three sites, three products, one deficiency. In each case the conclusion arrived before the evidence, and the file could not show otherwise.

Phase I: What Has To Exist Before You Invalidate Anything

A laboratory investigation is not a search for a reason to discard the number. It is a test of a stated hypothesis, and the order matters: hypothesis first, testing second, conclusion last.²

Four things belong in the record before an initial result can be set aside:

• a written hypothesis of what went wrong, specific enough to be falsified
• evidence of an assignable cause, not the plausibility of one
• the checklist that the analyst actually completed, rather than the blank version filed with the procedure
• a review by the analyst and the supervisor performed while the work is fresh, not reconstructed weeks later when the batch decision is due.

“The analyst probably made an error” is a hypothesis. It becomes a finding only when something in the record supports it: a torn column frit, a mis-set wavelength, a diluent prepared from the wrong stock. Absent that, the result stands, and the investigation moves on.

The April 2026 letter is the clean illustration. A species that cannot be identified is not evidence of laboratory contamination. It is an open question, and the firm closed it by assumption.³

The Decision Point Into Phase II

This is the weakest link in most records, because it is the point at which the laboratory hands the problem to the quality unit and the paperwork changes hands with it.

Three questions have to be answerable from the file alone:

• what triggered the escalation, and when
• which testing plan was approved, by whom, and before any retesting began
• how the sample was controlled in between, including who held it and what it was subjected to.

The quality unit does not discharge its role by approving the plan and stepping back. If the retests pass and no assignable cause was ever found, the investigation does not end in a pass. It ends in a batch decision that has to be argued in writing, against the live possibility that the original result was telling the truth. That paragraph — the one explaining why the firm believes the first number did not represent the batch — is the paragraph most often missing from the file.

The testing plan is the part most often written after the fact. It should specify the number of retests, the analyst performing them, the acceptance criteria, and, crucially, what the firm will do if the retests pass. A plan that only describes the passing path is not a plan; it is an expectation. The South Korean case shows what happens when the plan is built to produce a particular answer: the retest becomes a rehearsal rather than a test.⁴

Averaging, Outliers, And Resampling

These are the three places where a failing result quietly disappears.

Averaging is legitimate when the procedure defines the average as the reportable result, and it is illegitimate when it is used to dilute a failing value with passing ones. If averaging conceals variability that the specification exists to detect, the average is not the result; it is a summary of results the inspector will ask to see individually.²

Outlier tests have a narrow role. They may support an investigation into microbiological assays, where variability is expected, but they cannot invalidate a chemical assay result on statistical grounds alone, and they cannot be applied retrospectively. If the test is not named in the procedure before the result appears, it is not available afterward.²

Resampling is the most abused of the three. It is defensible only when the original sample is demonstrably unsuitable — wrong container, compromised integrity, documented handling error. It is not defensible as a second opinion. FDA made the microbiological version of this argument explicitly in April 2026: contamination may not be uniformly distributed, so a passing second sample says nothing about the batch the first sample came from.³

The OOS Record As An ALCOA+ Chain

Laboratory records must contain complete data derived from all tests performed.⁶ Complete is the operative word, and it is what turns an OOS file into a data integrity finding rather than an investigation finding.

Attributable, legible, contemporaneous, original, accurate — with complete, consistent, enduring, and available added by common practice — is not a poster for the laboratory wall. It is the list of properties that lets a third party rebuild your decision.⁷ Attributable answers who decided. Contemporaneous answers when, and it is the property that collapses first: a review signed three weeks after the batch shipped tells the reader the sequence of events, whatever the conclusion says. Original answers what the instrument actually produced, before anyone transcribed it into a summary table.

The European case in May 2026 turned on exactly this: the investigations were performed, but the data behind them were not recorded as the work happened, and the firm could not show which conclusions rested on which observations.⁵

Apply one test to your own files. Two years from now, a reviewer who has never met your analysts opens this record. Can that person reconstruct what was tested, what was hypothesized, what was found, who decided, and when — without asking anyone a single question? If the answer depends on institutional memory, the file is incomplete today.

A Reporting Structure You Can Lift

The structure below carries the evidence an inspector asks for, in the order the questions come. Each heading is followed by what has to sit underneath it.

1. Event description — the result, the specification, the method, the analyst, the date and time.
2. Immediate assessment — what was checked before anything was touched, including system suitability and the integrity of the sample and standards.
3. Hypotheses tested — each hypothesis, the test performed against it, and the outcome, including the hypotheses that failed.
4. Assignable cause — the evidence, or an explicit statement that no assignable cause was found. The second is a legitimate outcome and a stronger record than a weak first.
5. Phase II plan — the approved plan, the approval date and signature, and the predefined treatment of results.
6. Results and statistical treatment — every individual value, with any averaging or outlier decision justified against the procedure that authorized it.
7. Batch impact and scope — other batches, other products, and other analysts potentially affected, and the basis for the boundary drawn.
8. Quality unit decision — the disposition, who made it, and the reasoning in one paragraph a nonspecialist can follow.
9. Corrective and preventive action (CAPA) and effectiveness — what changed and how the firm will know it worked.

None of this requires a new system. It requires the reasoning to be written at the moment it happens, by the person who did it, and kept where the next reader will find it. That is the whole of the discipline, and it is the difference between an investigation an inspector reads and an investigation an inspector cites.

References:

1. 21 CFR 211.192, Production record review. U.S. Government Publishing Office, Electronic Code of Federal Regulations.
2. Investigating Out-of-Specification (OOS) Test Results for Pharmaceutical Production — Level 2 Revision, Guidance for Industry, U.S. Food and Drug Administration, May 2022. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/investigating-out-specification-oos-test-results-pharmaceutical-production-level-2-revision
3. Warning Letter, Medical Products Laboratories, Inc., CMS #721916, U.S. Food and Drug Administration, April 9, 2026. https://www.fda.gov/inspections-compliance-enforcement-and-criminal-investigations/warning-letters/medical-products-laboratories-inc-721916-04092026
4. Warning Letter, Huons Co., Ltd., CMS #724650, U.S. Food and Drug Administration, June 15, 2026. https://www.fda.gov/inspections-compliance-enforcement-and-criminal-investigations/warning-letters/huons-co-ltd-724650-06152026
5. Warning Letter, Pharmathen International S.A., CMS #723680, U.S. Food and Drug Administration, May 27, 2026. https://www.fda.gov/inspections-compliance-enforcement-and-criminal-investigations/warning-letters/pharmathen-international-sa-723680-05272026
6. 21 CFR 211.194(a), Laboratory records. U.S. Government Publishing Office, Electronic Code of Federal Regulations.
7. Data Integrity and Compliance With Drug CGMP: Questions and Answers, Guidance for Industry, U.S. Food and Drug Administration. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/data-integrity-and-compliance-drug-cgmp-questions-and-answers

About The Author:

Aldo Xhango is the founder of GuideGxP, an independent publisher of operational GMP/GxP guides and standard operating procedures for pharmaceutical quality teams. His work is published at www.guidegxp.com, and he can be reached at aldo@guidegxp.com.