The Rising Tide: Why RECOVER's Brain Fog Trial Couldn't Find What Wasn't Lost

RECOVER-NEURO enrolled 328 adults with Long COVID cognitive symptoms across 22 U.S. sites. It tested five approaches to brain fog: computerized brain training (BrainHQ), BrainHQ plus structured cognitive rehabilitation (PASC-CoRE), BrainHQ plus active transcranial direct current stimulation (tDCS), BrainHQ plus sham tDCS, and an active comparator of crossword puzzles and games.

Every participant trained five days a week for ten weeks. Retention exceeded 90%. Adherence exceeded 80%. The trial ran exactly as designed.

Every comparison was null.

The Numbers

Comparison	Adj. Difference	95% CI	P
BrainHQ vs. puzzles	0.0	−0.2 to 0.2	0.98
PASC-CoRE + BrainHQ vs. puzzles	0.1	−0.1 to 0.3	0.18
Active tDCS vs. sham tDCS	0.0	−0.2 to 0.2	0.97
PASC-CoRE + BrainHQ vs. BrainHQ alone	0.1	−0.1 to 0.3	0.18

The target effect size was 0.5 on the modified ECog2, a 41-item self-report scale of everyday cognition. The confidence intervals are tight enough to exclude any clinically meaningful difference. These are not underpowered nulls. These are definitive.

The Tide

Here is the finding that matters more than the nulls: everyone got better.

Uniform improvement across all arms. Secondary measures confirm: PROMIS Cognitive scores rose 4.4–6.9 T-score points, Symbol Digit Modalities improved by 3.2–5.4 correct answers, Digit Vigilance reaction time dropped 16–37 milliseconds. Every metric moved in the right direction. None moved differently by arm.

The authors note that time since infection predicted improvement more than treatment assignment did. The tide was rising. All five interventions rode it. None caused it.

Three Design Failures

1. The Wrong Patients

RECOVER-NEURO enrolled on subjective cognitive complaints. The authors are explicit: "Low scores on objective cognitive tests were not required for participation."

The result: 60.9% of participants — 176 of 328 — had no objective cognitive impairment on neuropsychological testing. Their scores were within 1.5 standard deviations of population norms. Only 8.7% (25 patients) showed deficits across two or more cognitive domains.

A cognitive rehabilitation trial where six in ten participants have no measurable cognitive deficit is not testing cognitive rehabilitation. It is testing whether attention from a structured program makes people feel better. It does. Equally, regardless of what the program contains.

2. The Subjective-Objective Gap

This is not a criticism of the patients. The subjective-objective gap in Long COVID cognition is well-documented and likely reflects real neuroinflammatory processes that impair the experience of thinking without always producing measurable performance deficits on timed tests. But it means the trial's instrument was measuring something real — how these patients feel about their cognition — while its design assumed it was measuring something else: whether specific interventions change cognitive performance.

3. Exercise in a PEM Population

This is the sharpest finding in the entire trial, and the authors buried it.

65.2% of RECOVER-NEURO participants — 214 of 328 — reported post-exertional malaise at baseline. Two-thirds of this cognitive exercise trial's population crashes after exertion.

PEM was never analyzed as a treatment moderator. The trial protocol did not stratify by PEM status. The paper reports PEM prevalence at baseline and endpoint (44.7% at end, decreased across all arms) but does not ask the obvious question: did PEM patients respond differently?

The implied answer comes from a different RECOVER trial. RECOVER-ENERGIZE, testing exercise interventions for Long COVID fatigue, now screens PEM patients out of its exercise arms. The program learned the lesson. It just never said so about RECOVER-NEURO.

The Matched Evidence

This is the second consecutive RECOVER trial to fail in exactly the same way.

Same funding stream. Same umbrella condition. Different clinical domains. Identical failure mode. Both trials hit their mechanical targets — the interventions did what they were designed to do pharmacologically or cognitively — and both missed the patients. The drugs and therapies worked. The denominator didn't.

As Diaphorai has documented, the credibility crisis in modern research runs deeper than failed replications. These RECOVER trials may be perfectly reproducible — run them again, get the same null results. But they are epistemically void. They cannot answer the question they were designed to answer because the population labeled "Long COVID" contains mechanistically distinct conditions that respond to different interventions. The treatment signal drowns in the denominator noise.

The Three Real Questions

RECOVER-NEURO asked one question: does cognitive rehabilitation improve brain fog in Long COVID? That question assumes a single population with a single mechanism.

The data suggests at least three populations were enrolled:

1. The orexin-depleted. Patients whose brain fog stems from COVID's destruction of hypothalamic wake-promoting neurons (Post #17). Cognitive training cannot restore dead neurons. These patients need pharmacological orexin support.
2. The microclot-hypoperfused. Patients whose cognitive symptoms reflect reduced cerebral blood flow from persistent microclots and endothelial damage (Post #14, Post #15). Training a brain that isn't getting enough oxygen is exercise on a broken treadmill.
3. The naturally recovering. Patients whose symptoms were trending toward resolution regardless of intervention — the ones generating the rising tide. They would have improved with crossword puzzles, with BrainHQ, or with nothing.

Mix these three populations in the same trial and the treatment signal — if one exists for any subgroup — disappears into the average. The trial can't fail in an informative way because both failure and success are consistent with the category being wrong.

The $1.15 Billion Question

RECOVER has spent $1.15 billion. Two completed intervention trials. Two nulls. The program's organizational structure — its enrolled cohorts, its site infrastructure, its data systems — is built on the premise that "Long COVID" is a coherent category that can be treated as a unified population.

Admitting the category is wrong means admitting the trial infrastructure was wrong. Which means the funding structure was wrong. Each layer adds weight to the status quo. This is what taxonomic inertia looks like in real time: a $1.15 billion program that generates reproducible nulls because the category it was built to treat may not exist as a single treatable entity.

The remaining RECOVER trials — RECOVER-SLEEP, RECOVER-ENERGIZE, RECOVER-TLC — face the same denominator problem. Baricitinib may work on the JAK-STAT-driven inflammatory subgroup. Low-dose naltrexone may help the neuroinflammatory subset. Semaglutide may benefit the metabolic phenotype. But if all three are enrolled in the same "Long COVID" bucket, the signal may drown again.

RECOVER-ENERGIZE's decision to screen out PEM patients from exercise arms suggests the field is starting to learn. But screening out is not the same as stratifying in. The question isn't whether to exclude PEM patients — it's whether to design trials around mechanistically defined subgroups rather than a syndromic label.

Until that happens, we may keep running well-executed trials that produce well-powered nulls, while patients whose specific biology would respond to specific interventions are averaged out of existence.

---

Sources: Knopman et al., "Evaluation of Interventions for Cognitive Symptoms in Long COVID," JAMA Neurology 2026;83(1):49-59. RECOVER-AUTONOMIC ivabradine results from ACC.26. Charlton et al., "Post-exertional malaise and the myth of cardiac deconditioning," BJSM 2026. Appelman et al., "Exercise pathophysiology in long COVID," Nature Communications 2024. Taxonomic inertia framework via Diaphorai. This is Post #33. It connects to Start Here (Post #32), RECOVER-AUTONOMIC (Post #23), and The Treatment Graveyard (Post #8).