FDA Warning Letters Show What an Autonomous Lab Must Record
· Updated

An autonomous lab should keep every run it tries, including the ones that fail, together with the sample, controls, instrument and timing behind each result. That context is exactly what FDA found missing in 154 of the 540 drug manufacturing (CGMP) warning letters it has posted since January 2021.
The 2023 letter to Intas Pharmaceuticals shows what a missing record looks like. FDA found that an analyst "did not report all the test results", that balance printouts were "discarded in the trash", and that the site "aborted hundreds of chromatographic sequences" over almost three years (letter). The record held only the runs the lab chose to keep.
An autonomous lab, which picks its own experiments, runs them and reads the results, can open the same gap much faster. It can repeat an assay in minutes with nobody watching, so its record has to be designed before its first experiment.
FDA cites missing lab context in a growing share of letters
The share of drug CGMP warning letters that cite a lab record missing data, or the context needed to read it, rose from 18% in 2021 to 38% in 2026.
540 drug CGMP warning letters, 22 January 2021 to 10 November 2026224 letters about lab records, each failure quoted by a model324 of 324 quotes on these cards found word for word in the letter154 letters with at least one of these gaps; a letter can sit in several rows
Open a kind of gap to see the letters behind its count. The highlight marks FDA's words about what the record lacked.
67 of 540 letters, newest first after the ones the post quotes
FDAYou invalidated multiple out-of-specification (OOS) results for (b)(4) USP without adequate scientific justification.
Read the letterFDAYour QU failed to perform investigations into OOS results obtained during trial injections.
Read the letterFDAA fter obtaining passing results, you concluded that 60% ± 5% humidity storage condition was the cause for the aspirin sub-potency and (b)(4) impurities without scientific justification, maintaining that the OOS results "may not represent condition of typical lot 18J024 consumer goods in the marketplace." At the time, your firm took no market action.
Read the letter
Every card is in the data (JSON), and every finding is in the CSV below.
Three failures run through most of the 154 letters, and in each a run happened but its result or its context never reached the record. The most common, in 67 of the 154, was a failing result set aside without a real investigation. Close behind were records that lacked the sample preparation, the instrument or the time (57 letters), and data that were generated and then left out or deleted (56). Ava Inc., for example, ran a test twice, and "the more favorable result was reported" (letter).
These letters concern drug quality control labs, which fall under 21 CFR Part 211, not research labs. They are still the most detailed public record of how lab data fail under inspection. An autonomous lab that reruns assays on its own could produce all three failures in minutes, with nobody watching.
FDA expects failed runs in the record too
The rules behind these citations are plain about failed runs. Section 211.194(a) asks for "complete data derived from all tests", and FDA's data integrity guidance says failing and suspect data belong in the record. FDA's guidance on out-of-specification results goes further and calls testing until a pass "unscientific and objectionable" (OOS guidance). For a system that can rerun an assay in minutes, that means every rerun is a result to keep, linked to the run it repeats.
A complete record also matters after the fact. When FDA finds a gap, it asks how far back the gap goes, and 84 of the 224 lab-record letters ask for a review of past data or released batches. A lab can answer that question only if it kept every run.
A partial record turns a failed run into an improvement
A record that drops a run's controls and timing can make a failed run look like an improvement, and a record that drops the run itself hides it. Everything in the example below is fictional. No physical experiment took place, and the example is an exercise in record design, not a validated method. A food-color solution is read on an optical reader, and the rules require a reference reading between 0.95 and 1.05 and a reading within 10 minutes.
| Run | Sample signal | Reference signal | Elapsed minutes | Context and disposition |
|---|---|---|---|---|
| A | 0.40 | 1.00 | 8 | Identity and calibration recorded; accept |
| B | 0.80 | 0.60 | 28 | Queue delay, reference check failed; stop |
| C | 0.41 | 1.00 | 8 | Calibration record unknown; hold |
| D | 0.42 | 1.00 | 8 | New aliquot, authorized after B and linked to it; accept, keep B |
Read without its controls and timestamps, run B looks like a doubling of the signal in run A. With them, B fails both rules and is stopped. Run C has no calibration record, so nothing supports accepting it. A system that kept only runs A and D would show a clean result and no trace of B.
The fields a run record should hold
Most research labs fall under neither Part 211 nor FDA's rules for good laboratory practice (GLP), because Section 58.3(d) excludes basic exploratory studies from GLP (21 CFR 58.3). An autonomous research lab therefore has to design its own record. The fields below map to the parts of 211.194, although no rule requires this exact set and not every instrument can capture all of it.
| Component | Keep | 211.194 counterpart |
|---|---|---|
| Sample lineage | Source and aliquot identity, moves, preparation time | (a)(1), (a)(3) |
| Reagents | Identity, lot, storage history | (c) |
| Timing | Timestamps with time zone, queue delays | (a)(1), (a)(7) |
| Controls | Raw control results, the limits and rule version | 211.160(b) |
| Instrument state | Instrument ID, method version, calibration, errors | (a)(2), (d) |
| Raw result | Original output, units, processing version | (a)(4), (a)(5) |
| Disposition | What went wrong, who recorded it, the review decision, any linked rerun | (a)(8), 211.192 |
Two habits make these fields trustworthy. The first is to store each observation apart from any model's reading of it, and to record when a cue was not collected, since a missing observation logged as "normal" hides the gap. The second is to check that each field rests on something real: a timestamp needs a known clock, and a file ID needs a file a reviewer can open.
What we would check before trusting the results
- Every run is kept, and each rerun links to the run it follows.
- The system cannot delete or overwrite a raw result, and its audit trail is on.
- Observations are stored apart from interpretation.
- A scientist can rebuild one accepted run and one stopped run from the files alone.
Data and method
We read the 540 drug CGMP warning letters FDA posted from 22 January 2021 to 10 November 2026. A language model, Claude Opus 5.5, read the 224 letters about lab records and quoted each failure, and we confirmed all 1,109 quotes against the letter text. In a hand check of 24 letters, 128 of 134 findings had the right type and 7 failures were missed, so the counts are a floor. FDA posts some letters late, so the 2026 count will grow. The fictional table comes from a short program that checks each run against the example's invented rules.
Downloads: counts by year (CSV), every coded finding with FDA's words (CSV) and the fictional example (CSV).
Frequently asked questions
Does 21 CFR Part 211 apply to an autonomous research lab?
No. Part 211 covers drug manufacturing quality control, and GLP excludes basic exploratory studies. The letters are still the best public guide to how lab records fail.
Can a passing rerun replace a failed run?
Not without a real cause. FDA's OOS guidance says a failing result stands unless a lab or calculation error is found.