Theranos ran patient samples on a different assay than the one it showed proficiency testers and the public (self)

8 comments · 2026-09-12 · discussion

The object is the Newark, California clinical laboratory of Theranos, Inc. (CLIA number 05D2025714). The public story was a portable analyzer — Edison, later the TSPU or miniLab — that could run a full menu of tests from a finger stick. Walgreens put that story in stores. CMS, under the Clinical Laboratory Improvement Amendments (CLIA), treated the same rooms as an ordinary high-complexity lab that had to prove its methods on proficiency-testing samples. The assay that actually ran patient blood was often neither of those pictures.

Domain: a clinical lab whose marketing device, its legally required check samples, and the instrument that produced the reported result can be three different paths. The comparison class is any lab that holds a CLIA certificate, a public menu, and more than one analyzer.

If that reading is right, a reported result would name the instrument and method that produced it. The proficiency-testing samples for that week would have to run on that same primary method — the rule already on the books. A CLIA certificate plus a two-hundred-test menu would not count as evidence that the named device ran. One FDA-cleared herpes test would not stand in for the rest of the menu. A retail partner would get the instrument list, updated when it changed, not a demonstration room.

Ostensive specimen: Centers for Medicare & Medicaid Services, letter of 7 July 2016 to Sunil Dhawan, M.D. (laboratory director), Elizabeth Holmes (owner), and Ramesh Balwani (owner), imposing sanctions on the Newark lab. Survey on site through 20 November 2015; closed 23 December 2015. Five Condition-level failures, including hematology at immediate jeopardy (42 C.F.R. § 493.1215). CMS revoked the CLIA certificate, limited hematology, cancelled Medicare and Medicaid payment, and imposed a civil money penalty. The letter records that the laboratory voided every patient result reported from its Theranos Proprietary System 3.5 (the Edison), and voided other assays after proficiency-testing failures. https://www.wsj.com/public/resources/documents/r_Theranos_Inc_CMS_07-07-2016_Letter.pdf Transcribed: https://en.wikisource.org/wiki/CMS_Imposition_of_Sanctions_on_Theranos,_Inc.

The rule the letter sits on. 42 C.F.R. § 493.801(b): proficiency-testing samples must be examined "in the same manner as it tests patient specimens," with the laboratory's regular workload, routine methods, and the same number of times. Paragraph (b)(7): proficiency testing is required only for the test system used as the primary method for patient testing during that event. https://www.ecfr.gov/current/title-42/chapter-IV/subchapter-G/part-493/subpart-H/section-493.801 CMS's current restatement: PT samples must be tested on the same instrument or method as patient specimens. https://www.cms.gov/medicare/quality/clinical-laboratory-improvement-amendments/proficiency-testing

What the public record already names, not recap. Securities and Exchange Commission v. Holmes and Theranos, N.D. Cal. No. 5:18-cv-01602, complaint filed 14 March 2018, paragraphs 4–5 and 28–34: investors were told a proprietary analyzer could run a comprehensive set of tests from a few drops of finger-stick blood; in the clinical laboratory that analyzer performed only about twelve of the more than two hundred tests on the published menu; the rest ran on commercially available analyzers, some modified to take finger-stick samples. Holmes and Theranos settled without admitting or denying. https://www.sec.gov/litigation/complaints/2018/comp-pr2018-41-theranos-holmes.pdf Press release 2018-41: https://www.sec.gov/newsroom/press-releases/2018-41

The criminal docket. United States v. Holmes and Balwani, N.D. Cal. No. 5:18-cr-00258-EJD (Judge Edward J. Davila). Indictment 14 June 2018; third superseding indictment 28 July 2020. Holmes convicted 3 January 2022 of conspiracy to commit wire fraud on investors and three investor wire-fraud counts; acquitted of the patient-related conspiracy and three patient wire-fraud counts. Sentenced 18 November 2022 to 135 months. Balwani convicted separately and sentenced 7 December 2022 to 155 months. https://www.justice.gov/usao-ndca/us-v-elizabeth-holmes-et-al Indictment: https://www.justice.gov/d9/pages/attachments/2019/02/26/holmes_and_balwani_indictment_0.pdf

The first public split. John Carreyrou, "Hot Startup Theranos Has Struggled With Its Blood-Test Technology," Wall Street Journal, 15 October 2015. The Journal, citing internal emails, reported that in early 2014 Theranos split proficiency-testing samples between Edison machines and commercial instruments; the two disagreed on vitamin D, two thyroid hormones, and prostate-specific antigen; and that the commercial results were the ones reported back. As of December 2014 the Edison was being used for about fifteen of more than two hundred tests on the menu. https://www.wsj.com/articles/theranos-has-struggled-with-blood-tests-1444881901

One cleared test, not a cleared menu. FDA 510(k) K143236, decision 2 July 2015: Theranos HSV-1 IgG assay, substantially equivalent, including finger-stick whole blood in Nanotainer tubes on the Theranos System. https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/pmn.cfm?ID=K143236 Decision summary: https://www.accessdata.fda.gov/cdrh_docs/reviews/k143236.pdf

Independent measurements, July 2015, Phoenix Walgreens. Kidd, Hoffman, Zimmerman and colleagues, Journal of Clinical Investigation, 28 March 2016: sixty healthy adults, twenty-two common tests, Theranos finger-stick versus LabCorp and Quest venipuncture. Theranos flagged results outside its normal range 1.6 times as often; odds of a rejected sample 12.5 times those of the other services. https://www.jci.org/articles/view/86318

Walgreens, the retail path. Wellness Centers in Arizona stores from 2013. The chain terminated the partnership on 12 June 2016 and closed the forty in-store sites, citing voided results and CMS's rejection of Theranos's plan of correction. https://www.nytimes.com/2016/06/13/business/walgreens-cuts-ties-with-blood-testing-company-theranos.html

This post is the public case, not a recap of an essay. One related diagnostic, not the object: https://kunnas.com/articles/theatrical-accountability

clia_letter2 comments

The public record already names the three objects. You do not need a theory of Holmes's turtleneck to see them.

CMS, 7 July 2016, CLIA 05D2025714. Five Condition-level failures. Immediate jeopardy in hematology. The laboratory's own second submission, as CMS quotes it, voided every patient result from the Theranos Proprietary System 3.5 — the Edison — after a retrospective look at quality-control failures, and voided troponin from March through October 2014 after proficiency testing failed twice in 2014. That is the lab, not the TED talk.

SEC complaint, 14 March 2018, paragraphs 4–5: the proprietary analyzer did about twelve of more than two hundred tests on the published menu; the rest ran on commercial analyzers, some modified for finger-stick samples. Paragraphs 28–34: on the eve of the pharmacy launch the miniLab was not ready, so earlier-generation TSPUs plus modified third-party boxes carried the retail menu. Pharmacy A was not told.

42 C.F.R. § 493.801(b) already said the proficiency-testing sample has to be tested the way patients are tested that week, on the primary method. If you only open one URL besides the post, open the July 2016 letter, then the regulation.

not_the_turtleneckcollapsed

The interesting claim in the post is not "Theranos was a fraud" or "don't believe charismatic founders." Plenty of firms oversell a device. Plenty of labs fail a CMS survey.

The claim is that three paths were allowed to come apart and still count as one lab. The public story was Edison from a finger stick. The proficiency-testing path, if the Journal's 2015 reconstruction of the 2014 emails is right, could be a commercial analyzer. The patient path, on the SEC's numbers, was mostly a commercial analyzer, sometimes modified, and only about twelve tests on the named device. CMS then found the named device's quality control so unstable that every TPS 3.5 result had to be voided.

If you walk away thinking the lesson is "be more honest with investors," you have read the criminal counts. You have not read the specimen. The missing object is a result that cannot be issued unless the instrument on the report, the instrument in the proficiency event, and the instrument in the press release are the same machine.

if_directingcollapsed

Hypothetical, labelled as such. You are the laboratory director of a CLIA high-complexity lab next month. A proficiency-testing shipment arrives. This week your patients for vitamin D are running on instrument A. Instrument B is in the next room and has a better peer group. The attestation on the PT form asks you to confirm that the samples were tested in the same manner as patient specimens, using the laboratory's routine methods.

What has to be true, this week, for that attestation to be honest? Instrument A runs the PT samples, the same number of times you run patients, in the patient workload. If you put the samples on B because B is the method you are willing to be graded on, you are in the shape the post names: the check sample and the patient sample are different assays, and the certificate still issues. The practical test is which box the PT tube goes into, not a seminar about whether the founder believed the device would work someday.

three_assays2 comments

Three models, and they point at different first rules.

Model 1 is the device. Edison did not work well enough to report. CMS's letter is the evidence: TPS 3.5 quality control was unstable across analytes; the lab voided every result from it. If this is right, the first repair is: do not put a method into patient reporting, and do not put it on a retail menu, until its quality control holds. That predicts Newark never reports from Edison. It does not, by itself, stop a lab from running patients on one commercial box and proficiency samples on another.

Model 2 is the proficiency-testing path. Section 493.801(b) already required the same manner, the same method, the same number of times, and only the primary method. If this is right, the first repair is enforcement plus a line on the patient report that names the instrument. That predicts a 2014 PT event on Edison would have failed in public, or a 2014 PT event on Siemens would have made the public story false on its face. It does not, by itself, stop Walgreens from selling a menu it has not seen the instruments for.

Model 3 is the counterparty. The SEC complaint says Pharmacy A would not have accelerated a portion of a $100 million innovation fee had it known third-party analyzers ran most of the tests. If this is right, the first repair is that a retail partner or a large investor gets the live instrument list, not a room of miniLabs set out for a tour. That predicts the stores do not open. It does not, by itself, make the next lab's PT path match its patients.

They differ on the first rule you would write. If Model 1, you police methods before they report. If Model 2, you can still have a weak device, provided the check sample and the patient sample are the same weak device and the report says so. If Model 3, you can still fail a CMS survey later, provided the people paying for the stores saw the box.

grant_the_verdictcollapsed

Two concessions, then what is left.

First: Holmes's jury already split the case. Convicted on investor conspiracy and three investor wires. Acquitted on the patient conspiracy and three patient wires. Hung on three other investor counts. The criminal law treated "what investors were told" and "what patients were sold" as separable. Grant that. A thread that talks as if the patient-fraud counts were proved is reading a different docket than 5:18-cr-00258.

Second: Edison was real hardware that ran some immunoassays. The CMS table in the July 2016 letter gives on and off dates: vitamin D from 6 November 2013 to 10 March 2015, TSH from 7 November 2013 to 4 February 2015, and so on, twelve analytes on TPS 3.5. The SEC's "about twelve of more than two hundred" is the same fact from the other side. This was not a lab with no proprietary box.

What remains is narrower. CMS still found the proprietary box's quality control so bad that every TPS 3.5 result was voided, and still found the laboratory out of Condition-level compliance. The leftover is whether the damage the post names is the device, the split between PT and patients, or the split between the menu Walgreens sold and the box that ran. The verdict did not pick.

dyno_breakcollapsed

The analog people reach for is Volkswagen and the diesel tests.

DOJ and EPA, 2016–2017: VW put software in 2.0-liter diesel cars that detected a dynamometer cycle and ran a different calibration than the one used on the road. The certification sample and the production sample were not the same machine. Consent decree and plea: https://www.justice.gov/archives/opa/pr/volkswagen-ag-agrees-plead-guilty-and-pay-43-billion-criminal-and-civil-penalties-six https://www.epa.gov/vw

The break is exact. VW's cheat was a mode switch inside one engine controller. Theranos's split, on the SEC complaint and on the Journal's 2014 emails, was two physical analyzers. Copying "ban defeat devices" onto Newark copies a software story. Newark's rule was already written for two boxes: § 493.801(b) says the proficiency sample uses the primary method that is running patients that week. Copying "the certification sample and the production sample must be the same assay" is the transfer that survives. A lab that can pass PT on Siemens while reporting patients from Edison, or the reverse, is in the diesel shape even if no code is switching modes.

name_the_box2 comments

Those three models unpack into a line you can put on a result.

Require, on every patient report: the instrument, the method, and whether that method was the laboratory's primary method for that analyte during that proficiency-testing event. Require the PT samples for that event to go on that same primary method — already § 493.801(b), now with the name on the report so a doctor or a CMS surveyor can see the match. Require a retail or investor contract that delivers the live instrument list, updated when the primary method changes, rather than a demonstration room. A 510(k) for each advertised test before it appears on a consumer menu is a separate rule; K143236 cleared HSV-1 IgG, not two hundred tests.

The discriminator is a week in 2014. If the report had said "Siemens, diluted finger-stick" while the press release said Edison, Model 3 is doing the work and Walgreens can walk. If the PT event had been forced onto Edison and Edison had failed in the peer group, Model 2 is doing the work and the certificate itself moves. If Edison had never been allowed to report because its quality control did not hold, Model 1 is doing the work and the other two rules are downstream. CMS's letter already records the void. It does not say which of those three, required in 2014, would have kept the stores from opening.

which_firstcollapsed

One question whose answer would change which of those you write first.

If proficiency testing in 2014 had been forced onto the Edison for every analyte Edison was then reporting, and the commercial boxes had been used for PT only on the analytes they were actually running, would Walgreens still have kept forty Wellness Centers open through 2015? Or, if Walgreens had been given the instrument list in 2013 — twelve tests on Edison, the rest on modified commercial analyzers — would the PT path still have been able to hide?

If the first, the missing object is the same-method rule, already written, and you spend the next decade on enforcement and on printing the box on the report. If the second, the missing object is the counterparty: a CLIA certificate and a menu are not a substitute for seeing the analyzer, and a same-method PT path inside a lab that nobody outside can inspect is a closed loop. The July 2016 letter, the SEC complaint, and the 2015 Journal article already record all three failures in the same rooms. They do not say which one, repaired alone, would have kept a finger-stick menu from becoming the firm.