VISTA·RESEARCH™
How it worksThe LibraryResearchHeritageAboutTalk to the Desk
Essay

The Day Before Unblinding

By Russ W. Rosenzweig · Vista Research · 17 November 2026 · Why I write with AI

The page is dated Monday, November 24, 2025. It is version 3.0 of a plan, the written rules for judging a drug trial. The next day the trial's database was locked and the data were unblinded. Until then, nobody running the trial was to know which patients had received the drug.

The patients were boys and young men with Duchenne muscular dystrophy. Their average age was 15, and most could no longer walk. Each was to be followed for twelve months. The last of them finished the blinded year on June 18, 2025. The first version of the plan is dated July 2022. The final one came five months after that last visit and one day before the lock.

It matters which page counts. Judged by the first plan, the Food and Drug Administration says, the trial missed the customary line for a win. Judged by the last, by the agency's own figure, it crossed. So one trial has two answers. What could the people revising the plan have known by November 24? A decision about the drug is due November 22, 2026.

I have a weakness for dates like that Monday. At Vista Research we freeze our tests. Before the evidence exists, we write down the date, the result that would prove us right and the result that would prove us wrong. I am fond of this habit. I have been known to mention it.

The idea for this essay originated on November 7, 2026, in a publicly filed complaint: Nkamga v. Capricor Therapeutics, in the federal court for the Southern District of California. Citing the FDA's briefing documents, it alleges that the final version of the plan was created one day before the data were unblinded. The company says the plan was finalized before unblinding and governs the results. I take no side.

I followed the question instead, and it led me to arithmetic. By the customary standard, one honest test of a drug that does nothing is fooled one time in twenty. Let a team try five ways and keep the best, and it can be fooled as often as one time in four or five. Nobody has to cheat for that to happen. The same arithmetic governs good news of the other kind: "no cases reported" weighs almost nothing until someone says how many patients were watched, and for how long.

One day

Duchenne, in the words of the FDA's own briefing, "affects about 1 in 3,500 boys." They typically lose the ability to walk by age 12 or 13. Median life expectancy has risen to about 30 years, and heart disease and breathing failure remain the leading causes of death.

Capricor Therapeutics is a publicly traded company. Its candidate, deramiocel, was tested in a trial called HOPE-3. The FDA describes the trial as randomized, double blind and placebo controlled. It compared drug and placebo over twelve months in each patient.

Two terms first. In a double-blind trial, nobody involved knows who received the drug until the data are unblinded. The statistical analysis plan, or SAP, is the written rulebook for judging the result. It names the measurement that counts most, the primary endpoint. It says how that is compared and what to do about a patient who drops out. The rulebook is worth something because it is written before anyone can see which choices favor the drug.

The FDA laid out its timeline in a briefing document for its advisory committee, a panel of outside experts, which met on July 29, 2026. The first version of the plan, version 1.1, is dated July 19, 2022, nearly four weeks after the first patient's first visit. The last patient finished the double-blind period on June 18, 2025. A new version followed on September 26, 2025. Version 3.0 is dated November 24, 2025. The database was locked and the data were unblinded on November 25.

Version 3.0, the FDA wrote, "was created/finalized one day prior to database lock/data unblinding." The agency also described that version as "not discussed and consequently not agreed upon."

Then come the numbers. A p-value is the probability of a result at least as extreme as the one observed, assuming the drug has no effect and the test's other assumptions hold. The customary line is 0.05, one in twenty, and a result must fall below it to count. For the trial's main measure, a score of upper limb function, there are three numbers to hold.

Under the first plan, by the FDA's analysis, the p-value was 0.24. That is about one in four. When the FDA itself applied version 3.0, with its new rules for missing data, its figure was 0.045, below the line. In the company's final report the figure was 0.029.

What moved them? Part of the answer, the FDA says, is two rules for handling missing data that the first plan did not have. One, for patients who left after an adverse event, was already in the plan by December 2023, while patients were still enrolling. The other, for a patient's data after a prohibited medication, came in the final version. Between them they touched two patients, both on placebo. Run the new analysis with the old missing-data rules, and the FDA's figure for the main measure is 0.21. With the new rules it is 0.045. And why does the company's 0.029 still differ from that? The FDA says the company's analysis kept a "slow progressor" term in its model that the agency reads version 3.0 as removing, and "The inclusion of this variable lowered the p value." So by the FDA's own figure, the late plan also crossed the line.

One trial now has two sentences written about it. The FDA's briefing says HOPE-3 "did not meet its pre-specified primary and secondary efficacy endpoints." A company release of March 12, 2026, as the complaint quotes it, says the trial "met its primary endpoint." The patients are the same. The plans differ.

Capricor's account deserves the same space. "Our results are governed by the final analysis plan, SAP version 3.0, which was finalized prior to unblinding," it said on July 27, 2026. The version the FDA leans on, version 1.1, was in the company's words "an unsigned incomplete internal draft which became obsolete with the addition of cohort B," and it "did not include content specifically requested by FDA."

The company's own briefing document says the plan "was revised alongside the evolving protocol through documented revisions, all finalized before unblinding." It says the estimated treatment effect "is consistent across all models." The company also notes that The Lancet, a medical journal, published the trial's results in July.

The FDA describes version 1.1 differently. In its account, that is the version on which the trial was designed and powered, prepared by an outside research firm under the company's own approved blinding plan.

I found no document that settles it. The FDA's briefing itself grants that the two sides "have different views on which version of the SAP is most appropriate to use."

Both sides give the same two dates, November 24 and November 25. The argument is over what the one day between them means. To the company, a plan finished before the lock was finished in time. Its briefing says a key change was judged on blinded data "by personnel without any access to treatment assignment." To the FDA's reviewers, the double-blind period had ended five months earlier. They note that hypersensitivity reactions occurred in 22 of 53 patients on the drug and 8 of 52 on placebo, a pattern that can hint at who received what. Changes made that late, they wrote, "are at risk of being data-driven regardless of the formal blinding status."

The complaint alleges that Capricor's shares fell $12.70, or 64 percent, to $7.00 on July 27, 2026, the day the briefing documents appeared. That is an allegation in a pending suit. In August the company said the FDA had extended its review after a major amendment to the application. The decision is now due November 22, 2026. It is pending as I write, and I will not guess at it.

Five ways to look

Why should one day matter so much? Because of what happens to luck when it gets more than one try.

Take a trial of 300 patients, half on a drug and half on placebo, and suppose the drug does nothing. The team wrote down one analysis before the data were unblinded. Call it a win when the drug beats placebo at the one-in-twenty level, counted one way only, with the drug ahead. Luck hands that team a win 5 percent of the time, whatever the size of the trial.

Now suppose the team looks five ways after the data arrive and reports the best. The main measure at twelve months. The same measure at six. The share of patients who improved by a set amount. Only the patients who took every dose. An adjustment for how sick each patient was at the start. Any of the five could be defended at a conference. Assume for the moment that the five looks are independent, so that no one of them says anything about the others. The drug that does nothing then wins at least once 22.6 percent of the time, more than one time in five. With ten looks it wins 40 percent of the time.

Five looks at the same patients are not independent. They tend to win or lose together, and how strongly is an assumption I can vary but cannot measure. If the looks overlap modestly, the false win arrives about 19 percent of the time. If they overlap heavily, about 12 percent. The figure never returns to 5 unless the five looks are one look.

0%10%20%30%40%12345678910Analyses tried after the data are seen, the best one reportedChance of a false winOne honest test, written down first: 5%40%31%25%15%Independent analyses: best of 5 gives 22.6%Correlation 0.3: best of 5 gives 19.3%Correlation 0.5: best of 5 gives 16.6%Correlation 0.8: best of 5 gives 11.7%
More looks, more false wins. When a drug does nothing, one test written down before the data are seen calls it a win 5 percent of the time. A team that tries several analyses after seeing the data and reports the best wins more often: with 5 independent analyses, about 22.6 percent of the time, and with 10, about 40 percent. Analyses of the same patients overlap, which slows the climb but never stops it; the correlations shown are assumptions, not measurements. A win here is the drug ahead of placebo with a one-sided p-value below 0.05. Illustrative arithmetic, computed exactly and checked by a simulation of 1,000,000 trials; not a finding about any trial.

One caution about the line itself. Regulators usually count the one in twenty both ways, for a drug that looks better and for one that looks worse, so a drug that does nothing wins by luck one time in forty. By that line one honest look is fooled 2.5 percent of the time and the best of five independent looks about 12 percent, so five looks multiply luck by nearly the same amount. And all of this is an illustration of how luck compounds. It measures nothing about any company, and it does not say that anyone looked five ways.

There is an honest price for five looks. The usual correction makes each one clear a line near one in a hundred. A trial that wants to keep an 80 percent chance of catching a real effect then needs about 62 percent more patients. In a disease that strikes one boy in 3,500, patients are the scarce thing.

The statistician Andrew Gelman and the psychologist Eric Loken borrowed a title from Borges for this problem: the garden of forking paths. Their point was that nobody has to set out to find a result. A researcher who makes one reasonable choice at each fork, after seeing the data, has still walked a path the data helped choose.

That is why the date is evidence. A plan fixed before anyone could see the data is the only proof that there was one path. It is also what the FDA and Capricor are arguing about: whether November 24 was still before.

A changed plan, and an unplanned look

A changed plan does not always flatter a drug. I own shares of Regeneron, so I read this one as a shareholder. Regeneron, a publicly traded company, tested its drug fianlimab in combination with Libtayo against pembrolizumab, another company's drug, in advanced melanoma, a skin cancer. A trial like this is read out once a set number of patients have seen their cancer grow or have died. Through 2025 those events came slowly. What happened next is the account of the company's senior vice president of investor relations, as quoted in the complaint in Cheatham v. Regeneron. A later cohort of 360 patients was meant to count only toward a second measure. Then, he said, "this was in the fall of 2025," the company decided to add the later patients to the primary analysis "without changing the number of events required to trigger the readout." In March 2026, by the complaint's quotation, he had told investors "we are totally blinded to the data."

The result came on May 15, 2026. The trial "did not reach statistical significance." By the company's account, then, the plan was changed while the company was blind to the data, and the drug missed anyway. The complaint's grievance is about what shareholders were told and when, and that belongs to the court.

Replimune, also publicly traded, met the question from the other side. It sought approval of a melanoma drug it called RP1, given together with another drug, nivolumab. The FDA answered on April 10, 2026 with a complete response letter, its form of not yet. The letter says the application's main support was a single-arm study of 140 patients with no concurrent control group. To support its resubmission, the letter says, the company provided data from "an early unplanned analysis" of its ongoing late-stage trial, called Phase 3: 22 patients in one arm and 18 in the other, "representing 10% of the planned enrollment of 400 patients." Among the deficiencies, the FDA listed difficulty interpreting the progression-free survival data "due to lack of pre-specification for this analysis and without adequate type 1 error control." Type 1 error is the statistician's name for the false win.

Replimune disagreed the same day. "Following an agency suggestion," it said, it had submitted a proposal for a descriptive analysis from that trial. Then, on August 6, 2026, the FDA granted the drug accelerated approval on the single-arm study, with a confirmatory trial required. An unplanned look at a tenth of a trial is weak evidence for a drug. It is no evidence against one.

No cases to date

"There have been no reported cases of neoplasia following Itvisma administration to date." Neoplasia is the medical word for tumor growth. The sentence is the FDA's, from a notice of November 6, 2026 titled "FDA Investigating Serious Risk of Malignancy Linked to Zolgensma Gene Therapy."

Zolgensma is a gene therapy for spinal muscular atrophy, given to children under 2. The agency "has received postmarketing adverse event reports of malignancies in patients treated with Zolgensma." A sister product, Itvisma, is given to patients 2 and older, and the FDA determined that the risk applies to both, "as the two products contain the same vector and construct."

The Itvisma sentence is a count, and the count is zero. That is a different thing from the sentence press releases favor. Regeneron's release of May 15 says: "No new safety signals were identified with the fianlimab combination." A statement like that is a judgment about the evidence. It can be true even when adverse events were reported, and I have no quarrel with it. A zero is plainer, and it can be weighed the way a win can.

The sentence after the FDA's zero gives the approval dates. Zolgensma, May 24, 2019. Itvisma, November 24, 2025.

November 24, 2025. That is the Monday on which version 3.0 of Capricor's plan is dated. It is a coincidence, and I will ask nothing of it except that it help me remember that a zero has a date too. Itvisma had been approved for 347 days when the FDA wrote its sentence. The notice gives no number of patients treated, no years of follow-up, and no time from treatment to tumor in the Zolgensma reports.

The FDA did not lean on its zero. It applied the risk to both products. It said that "The overall benefits of Zolgensma and Itvisma continue to outweigh their potential risks for their approved uses." And it noted that Novartis, "the sponsor, is conducting long-term follow-up safety studies." I make no claim about the risk of either product. My interest is in what a zero can weigh, and the figures that follow are illustrations with rates I chose.

Start with the count. Suppose 200 patients were each watched through the whole period in which a harm could appear, and none had it. The data still allow, with 95 percent confidence, a risk as high as about 1.5 in 100. Statisticians call the shortcut the rule of three: divide three by the number of patients. If only half of such harms would be noticed and reported, the same zero allows about 3 in 100.

Then the clock. Suppose a harm strikes at a steady rate of 1 patient in 1,000 each year, begins only after some latency, a delay between treatment and harm, and is noticed half the time. Watch 1,000 patients for two years past that latency, and the chance of seeing even one case is about 63 percent. Watch them for a fifth of a year, and it is under 10 percent. Stop watching before the latency ends, and the chance is zero. The report will say "no cases" whether the harm is real or not. To be 95 percent sure of seeing a single case, those 1,000 patients must be followed about six years past the latency.

The philosopher Carl Hempel left a puzzle that fits. "All ravens are black" says the same thing, in logic, as "everything that is not black is not a raven." So a white shoe, being neither black nor a raven, confirms it. It does, by a hair, and by nothing at all if you picked the shoe up knowing it was a shoe. A report of no cases from patients not yet followed past the latency is the white shoe. It is true, and it weighs nothing.

A win needs the date of its plan. A zero needs its count and its clock. In both, the missing piece is how many chances luck was given.

Both sides of the same loss

Return to the boys. In 2016 the FDA approved the first drug for Duchenne, eteplirsen, sold as Exondys 51, under its accelerated pathway. Its main clinical evidence was a "12-patient clinical study," in the FDA's words, and the approval rested on a rise in a muscle protein called dystrophin. The FDA's briefing memorandum had called its best estimate of that protein, "0.9% of normal after 3.5 years of treatment," disappointing. The agency's own announcement of the approval said: "A clinical benefit of Exondys 51, including improved motor function, has not been established."

In June 2025 the FDA said it "has received two reports of fatal acute liver failure" after treatment with a Duchenne gene therapy, Elevidys, in boys who could no longer walk. In November 2025 the FDA approved a boxed warning for it, the strongest kind, and its approved use no longer included those boys.

Within a decade, in one disease, the standard of evidence has moved in both directions, and children were on both sides of it.

I think both sides of this tragedy are right. A fifteen-year-old with Duchenne cannot easily wait for a second trial, and every month of rigor is paid for in muscle. A drug approved on a result that luck produced costs children too: a place in another trial, a family's hope, sometimes safety. The parent who says approve it and the reviewer who says not on this plan are standing on two sides of the same loss. I would not want either one's job.

What the registry did not show

I went looking for an industry, and I should say plainly that I did not find one.

We took a first look at ClinicalTrials.gov, the federal registry: 4,743 industry-sponsored trials in the late stages called Phase 3 or Phase 2/3, with posted results and a dated analysis plan. The registry does not record the day of unblinding, which is the one date this essay is about. So the first look could answer only a narrower question. Is the plan dated before or after the trial's primary completion date, when its main measurements are finished? And do the late-dated plans win more often?

Late dates turn out to be ordinary. A little over half the plans, 52.3 percent, are dated after primary completion, mostly within about three months, the usual window for finishing a plan. Of the 2,263 trials with early-dated plans, 1,029 reported a p-value for their main result, and 689 of those were below 0.05. That is 67.0 percent. Of the 2,480 with late-dated plans, 1,331 reported one, and 926 were below 0.05. That is 69.6 percent. The gap is 2.6 percentage points. Its 95 percent confidence interval runs from minus 1.2 points to plus 6.4, so the data cannot tell it from zero. Late-dated plans showed no bunching of results just under the one-in-twenty line.

That is a first look, and I would not call it more. I came back with three court files and a notice. They are examples of a question, each on its own. None of them is a verdict.

The same rule in other rooms

Economists who run field experiments now commonly file a pre-analysis plan before they collect the data. Technology companies that test one version of a page against another learned that peeking at a live test every morning, and stopping when it looks good, is five looks under another name.

And we freeze our tests, which is where I came in. Vista's Library of models keeps several entries for this family of mistakes. Three of them are "What result would prove us wrong?", "The story that fits too well" and "The white shoe and the black raven." This essay leaves me a footnote to all three. Fixing a test in advance is necessary. It still has to be a test a wrong idea would fail.

Four questions before a decision date

Someone who owns a biotechnology stock ahead of an FDA decision can ask four things. The answers are sometimes in the briefing documents the agency posts before an advisory meeting.

Was the analysis plan agreed with the FDA, or only sent to it? When is the final version dated, measured against the day of unblinding? Does the claim rest on a look the trial did not plan? When a company reports a zero, does the zero come with a count of patients and a length of follow-up?

The complaints against Capricor and Replimune each allege a one-day fall of 64 percent, so the stakes are plain. The questions are modest tools all the same. None of them predicts what the FDA will do. They tell you how much weight the number in a press release can bear.

The date on the page

There is more to learn. The registry measurement deserves to be done properly, sponsor by sponsor. The people who know the rest are trial statisticians, former FDA reviewers, biotech finance chiefs and parents. Likely, we can get some of it from primary research, which is what Vista is all about.

The FDA's decision on deramiocel is due November 22, 2026. Whatever it says, the file will still hold a plan dated November 24, 2025, a database locked and unblinded on November 25, and two accounts of what the day between them means.

A plan's date is evidence. Ours are too. Each test Vista freezes carries its date, and one day a reader will set that date beside the day the evidence came in. I do not know what any of our tests will show. I would like to go on not knowing until the day.

Sources

Disclosures

This essay is analysis and commentary, not investment advice, and nothing in it is a recommendation about any security. Its numerical examples are illustrations. Russ W. Rosenzweig, Vista's founder, owns shares of Regeneron and holds no options on it, and owns no shares of Capricor, Replimune or Novartis and holds no options on any of them. Holdings through mutual funds and exchange-traded funds are not counted. How this essay was made: written with the help of AI tools. The models were built and every model figure computed in code, each public source was checked against the copy named in Sources, and the draft was fact-checked before I read it. I read it and listened to it before it was published. Why I write with AI.