Based on: Bacteria and Viruses read through the Vacuum.Net Theory

Jump to the article here

Medicine is full of catalogues.

Diseases are sorted by organ. Or by pathogen. Or by mechanism.

But none of these sorting systems tells you what can actually be done.

This paper offers a different cut.

Every disease state can end in one of only four ways.

1. Let it close. The body finishes the job itself. No treatment needed.
2. Remove one party. Kill or clear the other participant. Antibiotics are the classic case.
3. Supply what is missing. Replace an absent component. Insulin is the classic case.
4. Restore proportion. Both parties stay. What is wrong is the relation between them. Nothing is removed. Nothing is supplied. The balance must change.

Three of these four operations are crowded. They have industries, trial designs, and regulators built around them. The fourth is almost untouched. Yet the fourth holds sepsis, ARDS, severe influenza, severe COVID-19, the autoimmune diseases, Ebola, and — on the evidence in the paper — atherosclerotic cardiovascular disease. That last one kills more people than anything else on earth.

The asymmetry is historical, not biological. Operation 2 produced the most spectacular results in the history of medicine. The whole field then reorganised around it: find the agent, kill the agent. Operation 3 followed. Genetic medicine has made it technically routine. Operation 4 never got a name. It never got a standard trial design. It never got a regulatory pathway. And that is exactly where the largest remaining burden of death now sits.

Three consequences follow, and the paper builds on them.

The operation determines the endpoint. An operation-2 trial asks: was the agent cleared? An operation-4 trial cannot ask that. Nothing is being cleared. Measuring clearance produces null results for drugs that work.

The operation determines who to enrol. Operations 1 to 3 enrol by diagnosis. Operation 4 cannot. The same diagnosis can contain patients on both sides of the line where a treatment reverses its effect.

The wrong operation is not neutral. It causes harm. Several trials below quantify that harm.

2. Operation 1: let it close

Some diseases are self-limiting. The body completes the process on its own. Mild influenza is one. Viral pneumonia is another. Most wounds heal by themselves. Intervention adds nothing here. It can subtract.

The research question is not a treatment. It is: where is the boundary? Treating a patient whose process is still completing causes measurable harm.

RECOVERY gave the number. Among hospitalised COVID-19 patients who needed no oxygen, 28-day mortality was 17.0% on dexamethasone against 13.2% with usual care. That is a rate ratio of 1.22. The drug made things worse in exactly the group whose disease was still resolving. The confidence interval includes 1, so this is a direction, not a proven fact. But the point estimate points 3.8 percentage points the wrong way.

Antibiotics for viral infections repeat the same error at a much larger scale. The bill for that arrives in the next section.

Research direction 1: build the boundary test, disease by disease. The right clinical question is not “does this drug work?” It is “has this patient’s process stopped completing?” Today, that question has a bedside answer in exactly one condition: COVID-19, where the answer is whether the patient needs oxygen. Even that was found by accident, in a subgroup analysis. For every other condition in this row, the boundary is unmeasured. Success would be a simple admission measurement. It should separate patients whose process will complete from those whose will not. It should be validated against outcome without treatment.

3. Operation 2: remove one party

This operation built modern medicine. It still works beautifully where it works at all. Tuberculosis is curable. So are malaria, bacterial pneumonia, and hepatitis C. But two facts should govern further investment.

It is running into a wall. Deaths directly attributable to bacterial antimicrobial resistance were 1.14 million in 2021. The forecast for 2050 is 1.91 million. That is a rise of 67.5%. Associated deaths rise from 4.71 million to 8.22 million, up 74.5%. Cumulatively, 39 million deaths are forecast between 2025 and 2050. Deaths in adults over 70 rose more than 80% between 1990 and 2021.

The wall is structural, not technical. The operation works by removing a self-replicating party. A self-replicating party selects for escape. Every success narrows the space of future successes. No other operation has this property.

Where it works, the drug is no longer the constraint. Hepatitis C is the cleanest case. About 47 million people live with chronic infection. Roughly 68 million were ever infected between 2015 and 2024. Only about 24 million knew their diagnosis — 35%. Only about 13 million had received the direct-acting antivirals by the end of 2024. That is 19% of the ever-infected. Around 44 million people were never diagnosed at all. A curative eight-week regimen exists. Four out of five people who need it have not had it. No new molecule closes that gap.

This leads to three directions:

• Direction 2: treat delivery as the research problem, not the molecule. Where a cure already exists, the remaining deaths are a diagnosis and logistics problem. That problem is bigger than any new compound.
• Direction 3: stop paying for removal where nothing needs removing. Antibiotic use in self-limiting viral illness drives the resistance curve above. The boundary test of Direction 1 is also the cheapest lever on antimicrobial resistance.
• Direction 4: for resistant organisms, change operation, not compound. Phage therapy, anti-virulence drugs, and host-directed approaches all abandon removal. The argument for them is not that they are promising. It is that removal has a mathematical ceiling, and these do not.

The same ceiling, inside one patient

The resistance curve above describes a population over decades. Oncology shows the identical structure inside one body, over months.

Osimertinib is the best first-line drug for EGFR-mutated lung cancer. In the FLAURA trial, median progression-free survival was 18.9 months against 10.2 months on the older agents. That is a hazard ratio of 0.46 — close to a doubling. Median overall survival was 38.6 months against 31.8 months. That is a gain of 6.8 months, a hazard ratio of 0.80. At three years, 28% of patients were still on the regimen. Which means 72% were not.

Look at the gap between those numbers. Time before return nearly doubles. Survival gains only a fifth. What is removed is removed. What remains is what the removal did not reach. And what remains keeps growing. This is the antimicrobial resistance curve, with one body replacing the population and eighteen months replacing the decades.

So cancer is not a second instance of operation 2 waiting for a better compound. Nothing foreign is present. Nothing is missing. What does not settle is division itself. That places cancer structurally in the fourth row. It is the one fourth-row condition with no established boundary marker. Naming that marker is the open problem. Finding the next inhibitor is not.

4. Operation 3: supply what is missing

Where a component is absent, supplying it ends the disease. Insulin and vitamin C are the historical cases. The modern form is genetic. And it now works.

Exagamglogene autotemcel (exa-cel) edits BCL11A to restore fetal haemoglobin. Across more than 100 treated patients — 46 with sickle cell disease, 56 with beta-thalassaemia — the results are categorical. In sickle cell disease, 92.3% of evaluable patients had no vaso-occlusive crisis for at least 12 consecutive months. The longest crisis-free period was 54.8 months. In thalassaemia, 94.2% became transfusion-independent for at least 12 consecutive months, the longest for 59.4 months. A single administration converted a lifelong disease into its absence. The effect is approaching five years old and still holding.

The constraint is not biology. Each therapy is built once, for one mutation. The cost can only be amortised over large patient populations. That excludes most of the conditions the operation could solve. There are thousands of single-gene disorders. Approved genetic therapies exist for only tens of them.

Direction 5: make the platform the product, not the therapy. The target is a delivery and editing platform where changing the target sequence does not restart development, manufacturing, or regulatory review. Judge every proposal by what it does to the cost of the second indication, not the first. Two developments bear on this. Targeted lipid nanoparticles can generate engineered cells inside the body, removing per-patient manufacturing. And computational protein design can now generate binders and antibodies against specified targets with atomic accuracy, shortening the pipeline’s front end.

Direction 6: audit the boundary between operations 3 and 4. Not every deficiency is a deficiency. When supplying the “missing” component fails to end the disease, the classification was wrong. The condition belongs in the fourth row. The audit is cheap. It uses trial results that already exist.

5. Operation 4: restore proportion

Here, both parties remain. Nothing is cleared. Nothing is missing. The relation between the parts is what is wrong. The reference catalogue marks these conditions “relief only”: nothing removes them. Severe influenza is filed that way. So are sepsis, ARDS, severe COVID-19, the autoimmune diseases, and Ebola virus disease — for which no licensed vaccine or therapeutic exists against the Bundibugyo species behind the 2026 outbreak in the Democratic Republic of the Congo.

The first assumption to break is inherited from operation 2: that more response is better. It is not.

A Sierra Leone cohort of 112 Ebola patients showed this clearly. Of 246 differentially abundant transcripts, 220 were higher in those who died. Their interferon signalling was stronger. Their acute-phase responses were stronger. Fibrinogen levels were roughly 54-fold, 28-fold, and 19-fold higher in fatal cases. Meanwhile, viral load failed to separate outcome where it mattered. At the extremes, cycle threshold predicted death (Ct 12) or survival (Ct 30). Around the cohort mean of 21.4, outcome was close to even. Host-gene classifiers reached 79–92% accuracy in exactly that range. In this row, the size of the response is not a measure of defence. It is a measure of what is not settling.

Why trials in this row keep returning null

Because the intervention reverses sign inside the enrolled population. A single pooled estimate averages the two halves. RECOVERY showed the reversal inside one randomisation:

Respiratory support at entry	Dexamethasone	Usual care	Rate ratio
None	17.0%	13.2%	1.22 (0.93–1.61) — harm
Oxygen only	21.5%	25.0%	0.80 (0.70–0.92) — benefit
Invasive ventilation	29.0%	40.7%	0.65 (0.51–0.82) — strong benefit

Mortality under usual care rises across the strata: 13.2, 25.0, 40.7. The rate ratio falls: 1.22, 0.80, 0.65. The same drug, at the same dose, in one trial, changes sign.

HARP-2 showed what happens without stratification. It randomised 540 ARDS patients to simvastatin or placebo and found no overall survival difference. A secondary analysis then applied latent class analysis to baseline data alone — no outcomes used. Two classes emerged: 353 patients (65%) hypoinflammatory, 186 (35%) hyperinflammatory. Their baseline 28-day mortality was 39% against 17%. Within the hyperinflammatory class, simvastatin significantly improved 28-day survival (P = 0.008). A trial filed as negative contained a benefit. The stratifier was computable from measurements taken before treatment.

SAVE-MORE showed the converse, by getting the design right from the start. Anakinra was given to hospitalised COVID-19 patients selected by plasma suPAR of 6 ng/mL or more, excluding the most severely hypoxic. The primary outcome improved with an odds ratio of 0.36. 28-day mortality was 3.2% against 6.9%. Entry was decided by a measured marker, not by clinical stage. The boundary is therefore not lateness and not crude severity. It is a state that can be measured. A patient can be past it while still looking mild.

Direction 7: stratify at randomisation on a measured marker, and report the strata separately. Every trial in this row should be designed this way. A single pooled estimate is not conservative here. It is a design that destroys the signal.

6. Operation 4 outside the intensive care unit

Everything above comes from critical care, where death arrives in days. If the fourth row were confined there, it would be a small category. It is not.

CANTOS tested canakinumab, an antibody against interleukin-1-beta, in 10,061 people who had had a myocardial infarction and whose baseline hsCRP was 2 mg/L or above. Entry was already decided by a measured marker, as in SAVE-MORE. Three results matter.

The lipid route was not touched. LDL cholesterol went slightly up on the 150 mg dose (+1.7%) and events went down. Fifty years of cardiology have aimed at that lipid quantity. Whatever changed, it was not that.

Pooled, the effect is real and small. Primary events ran at 3.86 per 100 person-years on 150 mg against 4.50 on placebo. That is a hazard ratio of 0.85 — a 14.2% reduction.

The statistical space was divided across the wrong thing. The trial split its alpha across three doses of one drug. The 150 mg dose passed its prespecified threshold by a margin of 0.0004. The 300 mg dose produced essentially the same effect (HR 0.86) but was declared non-significant, because its own threshold was stricter and its P value nearly three times that threshold. Two doses of one antibody produced the same effect. One counted. One did not.

Then the division that matters: on-treatment hsCRP at three months, below or above 2 mg/L.

Endpoint	hsCRP < 2 mg/L	hsCRP ≥ 2 mg/L
MI, stroke or cardiovascular death	0.75 (0.66–0.85)	0.90 (0.79–1.02)
Cardiovascular death	0.69 (0.56–0.85)	0.99 (0.82–1.21)
Death from any cause	0.69 (0.58–0.81)	1.05 (0.90–1.22)
Number needed to treat over 5 years	16	57

All-cause death falls by nearly a third on one side of the line and crosses 1 on the other. The NNT differs 3.6-fold. The share of patients reaching the lower state rose with dose: 44%, 55%, 65%.

An honest caveat: this division is weaker evidence than HARP-2 or SAVE-MORE. HARP-2 classified on baseline data alone. SAVE-MORE decided entry before randomisation. CANTOS divided on a measurement taken three months after treatment began. The groups are therefore defined partly by how they responded. A group that responds may simply have had more room to improve. What CANTOS establishes is not that this marker identifies a treatable stratum. It is that a pooled estimate can conceal a division large enough to reverse the sign of mortality. The trial that would settle it: enter on the baseline marker, dose, measure at three months, randomise again.

The harms are in the same record. Fatal infection or sepsis ran at 0.31 per 100 person-years against 0.18. Against that, fatal cancer fell (0.45 against 0.64) and the primary endpoint fell by 0.64. This is not a defence of the drug, which was never licensed for this indication. It is a description of the fourth row: nothing is removed, and the cost appears exactly where the capacity to settle an infection is reduced.

The consequence is the most important in the paper. Atherosclerotic cardiovascular disease is the largest single cause of death in the world. If it belongs in the fourth row — and the LDL result is the strongest available argument that it does — then the fourth row is not a residual category of rare and desperate conditions. It is the largest of the four. And it has no standard trial design, no accepted biomarker class, and no regulatory precedent.

Direction 10: divide the statistical space across strata, not across doses. CANTOS spent its alpha on three doses and had none left for the division that reversed mortality. The correction is purely procedural: prespecify fewer doses and more strata. The dose can be found afterwards. The stratum cannot be recovered once the result is filed as negative.

7. The first clean demonstration that operation 4 is real

Autoimmune disease is the purest case in the fourth row. Nothing is missing. Nothing foreign is present. The relation is wrong. And treatment has consisted of holding it down indefinitely.

Eight patients with treatment-resistant systemic lupus erythematosus received a single dose of CD19-directed CAR T cells between March 2021 and June 2023. All eight entered remission, with SLEDAI scores of zero. All came off immunosuppressants, including glucocorticoids. Autoantibodies disappeared — one antibody in one patient excepted — and stayed negative through 12 to 24 months of follow-up. Naive B cells returned within months. The disease did not return with them. Patients retained vaccine responses to measles, mumps, rubella, chickenpox, Epstein-Barr, tetanus, and pneumococcus.

The B cell compartment was not removed permanently. It was depleted. It re-formed. And what came back did not reproduce the disease. The relation reset.

The caveats are real. Eight patients. An ethnically narrow cohort. Cytokine release syndrome and neurotoxicity as risks. No US approval for this indication. A cost of roughly $375,000 to $425,000 per infusion.

But the result is categorical, not incremental. These patients are not better managed. They are off treatment, with no disease activity. Nothing in the first three operations predicts that outcome for lupus.

Direction 8: pursue reset rather than suppression across the fourth row. The question to carry from lupus into sepsis, ARDS, and severe viral disease is not which mediator to block. It is which compartment can be made to re-form in a settled state. Depletion followed by reconstitution is one mechanism. It is unlikely to be the only one.

The cost structure is also where the directions converge. Generating engineered cells inside the body, by targeted nanoparticle delivery rather than external manufacture, is the step that would move a reset from an eight-patient demonstration to a usable intervention. That is the same platform problem as Direction 5. Solving it once serves both rows.

8. What it costs to pick the wrong operation

Alzheimer’s disease has absorbed more investment than any other fourth-row target — under an operation-2 programme: identify the agent, remove the agent. Amyloid was the agent. The removal works.

The clinical result does not follow. In the lecanemab phase 3 trial over 18 months, CDR-SB declined 1.21 points on treatment against 1.66 on placebo. That is a difference of 0.45 points, reported as “27% slowing”. For donanemab over 76 weeks, iADRS declined 10.2 against 13.1 — a difference of 3.4 points, reported as “33%”.

Now set those against the minimal clinically important differences: 1.0 on CDR-SB in mild cognitive impairment; 1.6 in mild dementia; 5 and 9 on iADRS. The lecanemab effect is 0.45 of the smaller threshold. One analysis concludes that changes of about 0.5 on CDR-SB and about 3 on iADRS are imperceptible to patients, families, and physicians.

The harms are not imperceptible. Nearly one in four lecanemab patients developed brain swelling or bleeding. More than one in three on donanemab. Treatment-related adverse events ran at 45% and 89%. Severe bleeding occurred in 5 patients against 1 on placebo for lecanemab, and 7 against 2 for donanemab. The number needed to harm is about 3 for both drugs. The number needed to treat could not be computed from the published data.

This is what applying operation 2 to a fourth-row condition produces. The agent is successfully removed. The relation is unchanged. The benefit sits below the threshold of perception. The harm does not.

The correct reading is not that amyloid is irrelevant. It is that removal was the wrong operation — for the same reason killing the virus is the wrong operation in severe influenza. The research question becomes: which relation in the brain has stopped settling, and what would make it settle? That is a fourth-row question. It has never been the one the field was funded to ask.

Direction 9: re-classify the major chronic diseases by operation before funding another removal programme. Neurodegeneration, metabolic disease, chronic kidney disease, heart failure, and the major psychiatric disorders have all been pursued predominantly as operation-2 or operation-3 problems. Each should be re-examined against the four rows. The trial results that already exist are sufficient to do it.

9. The ten directions, collected

#	Direction	Judged by
1	Build the boundary test, disease by disease	Prospective separation of untreated outcomes
2	Treat delivery as the research problem where a cure exists	Share of the ever-infected treated
3	Stop applying removal where nothing needs removing	Reduction in inappropriate antimicrobial courses
4	For resistant organisms and returning tumours, change operation, not compound	Durability under selection, not initial potency
5	Make the editing platform the product	Cost of the second indication
6	Audit the operation-3 / operation-4 boundary	Programmes correctly redirected
7	Stratify at randomisation on a measured marker	Sign reversal detected rather than averaged away
8	Pursue reset rather than suppression	Drug-free remission, not disease control
9	Re-classify the major chronic diseases by operation	Removal programmes stopped where the row is wrong
10	Divide the statistical space across strata, not across doses	Effects retained instead of filed as negative

Three of these — 2, 3, and 6 — require no new science. They are re-allocations, and they act on larger numbers than anything in the pipeline. Two — 5 and 8 — converge on one engineering problem: generating engineered cells inside the body. And two — 7 and 10 — are free. They are changes in how a trial is written down.

10. What follows for each reader

For clinicians. Before reaching for an intervention, name the operation. Where the process is still settling, suppression is not neutral. In RECOVERY’s no-oxygen group, the point estimate ran 3.8 percentage points against the patient. Where an operation-2 cure exists, the question that decides mortality is whether the patient was ever diagnosed — not which regimen to choose.

For trialists. A null result in the fourth row is not evidence of no effect. It is evidence of an unstratified population, until a stratified analysis says otherwise. Fix the strata before allocation. Choose the stratifier as a measured marker, not a clinical stage. Publish the strata separately. The archive of null immunomodulation trials is a research resource, not a closed file.

For funders. Three directions require no new science and act on the largest numbers. Two more cost nothing but a change in protocol writing. These compete poorly for funding because they produce no molecule. They are where the available mortality reduction is concentrated.

For industry. The fourth row is the largest unserved category in medicine. It is unserved because it has no standard trial design, no accepted biomarker class, and no regulatory precedent for a stratified claim. The company that establishes that pathway — a marker, a stratified design, a label naming a stratum rather than a diagnosis — defines the category. The commercial argument for the fourth row is not that it is virtuous. It is that it is where the volume is, and where no one has a design.

One caution. This framing does not predict that any particular intervention will work. It predicts where to look, which populations to enrol, and which negative results deserve re-examination. Every claim it makes is refutable by the trials it proposes.

11. Verification and limitations

The paper states that every figure was recomputed from the reported counts or percentages before being quoted, and it publishes the checks. Two discrepancies are left standing rather than smoothed: a hepatitis C treatment share that computes to 54% of the diagnosed where the source says about 20%; and a diagnosed share of 35% against a stated 36%, attributed to rounding.

The stated limitations are substantial and worth repeating. The CAR-T lupus result rests on eight patients in a single centre, with 12 to 24 months of follow-up, in an unapproved indication. The HARP-2 subphenotype analysis is secondary. The RECOVERY no-oxygen stratum’s confidence interval includes 1, so harm there is a direction, not a finding. The CANTOS division uses a post-treatment measurement, a weaker design than the other two trials. The FLAURA comparator allowed crossover, and compared against first-generation inhibitors rather than against no removal at all. The classification was checked in detail against one chapter of one reference catalogue, not exhaustively. Cancer’s placement in the fourth row is an argument from the shape of its failures, not a demonstration — no boundary marker for it is proposed. And the four-operation scheme itself derives nothing and predicts no effect size: what is refutable is every row assignment in it, and that is where it should be attacked.

---

Annotated reference list

The sources below are as cited in the original paper. The annotations summarise what each source contributed to the argument.

Stratified and biomarker-guided trials

1. RECOVERY Collaborative Group — dexamethasone in hospitalised COVID-19 (with subgroup outcomes as summarised by The Bottom Line and First10EM). The platform trial that provided the central sign-reversal evidence: rate ratios of 1.22, 0.80, and 0.65 across the no-oxygen, oxygen, and ventilated strata. It grounds the claims that operation determines benefit, that the wrong operation harms, and that a simple boundary marker (need for oxygen) already exists in one disease.
2. Calfee, C. S. et al. — ARDS subphenotypes and differential response to simvastatin (HARP-2 secondary analysis). Latent class analysis of 540 ARDS patients on baseline data alone. It split them into hyperinflammatory and hypoinflammatory classes with 39% versus 17% baseline mortality, and showed a significant simvastatin survival benefit in one class only. It is the paper’s demonstration that a “negative” trial can contain a real, recoverable effect.
3. SAVE-MORE (Kyriazopoulou et al.) — early anakinra guided by plasma suPAR in hospitalised COVID-19. A phase 3 trial that enrolled patients on a measured biomarker threshold (suPAR ≥ 6 ng/mL) rather than clinical stage. Odds ratio 0.36 on the primary outcome; 28-day mortality 3.2% versus 6.9%. It is the prospective proof that biomarker-guided entry makes an immunomodulator work.
4. RECOVERY — tocilizumab (summaries by REBEL EM and the European Respiratory Society); CAPE COD (NEJM 2023, Critical Care Reviews) and CAPE COD versus ESCAPE. Further critical-care immunomodulation trials used to establish the fourth row in the ICU: what worked, what failed, and how design choices — stratification, timing, and endpoint — decided the outcome.

Ebola virus disease

5. Liu et al., Genome Biology 2017 — Ebola patient transcriptomes. The Sierra Leone cohort of 112 patients. Source of the paper’s key conceptual break: 220 of 246 differentially abundant transcripts higher in fatal cases; fibrinogen fold-changes of ~54×, 28×, and 19×; viral load failing to predict outcome near the cohort mean while host-gene classifiers reached 79–92% accuracy. It shows that response magnitude is a measure of non-settling, not defence.
6. ECDC — multi-country Ebola outbreak 2026; LSHTM on Bundibugyo virus. Establishes that no licensed vaccine or therapeutic exists against the Bundibugyo virus species behind the 2026 Democratic Republic of the Congo outbreak — evidence that an entire lethal disease sits in the unserved fourth row.

Cell therapy and autoimmunity

7. American College of Rheumatology — CD19 CAR T-cell therapy in refractory systemic lupus erythematosus. The eight-patient series (treated March 2021–June 2023). All achieved SLEDAI-zero remission, discontinued immunosuppressants, lost autoantibodies, reconstituted naive B cells without relapse, and retained vaccine responses. It is the paper’s first clean demonstration that restoring proportion — reset rather than suppression — can end a fourth-row disease.

Gene editing

8. Vertex Pharmaceuticals — long-term exagamglogene autotemcel (exa-cel) data in sickle cell disease and beta-thalassaemia. The 46 + 56 patient dataset: 92.3% and 94.2% sustained crisis-free and transfusion-free rates at 12+ months, with the longest periods at 54.8 and 59.4 months and stable edited-allele levels. It shows operation 3 is technically solved — and, by its per-patient cost structure, why it is economically blocked.

Antimicrobial resistance

9. GRAM Project (University of Oxford) — global AMR burden forecasts. Source of the resistance ceiling figures: 1.14 million attributable deaths in 2021 rising to 1.91 million in 2050; associated deaths rising from 4.71 to 8.22 million; 39 million cumulative deaths projected 2025–2050; deaths in adults over 70 up more than 80% since 1990. It grounds the argument that operation 2 has a structural, mathematical limit.

Hepatitis C

10. “Working towards hepatitis C elimination” (Contagion Live) — global burden, cascade of care, and WHO 2030 targets. Source of the delivery-gap figures: ~47 million chronically infected, ~68 million ever infected, ~24 million diagnosed, ~13 million treated, ~44 million never diagnosed. It shows that where a cure exists, the binding constraint is diagnosis and logistics — not pharmacology.

Alzheimer’s disease

11. “Lecanemab and donanemab: an illustrated perspective on the data” (eNeuro, 2024). Source of the anti-amyloid analysis: the 0.45-point CDR-SB and 3.4-point iADRS differences, the minimal clinically important difference thresholds (1.0–1.6 and 5–9), the imperceptibility analysis (~0.5 and ~3 points), and the harms — brain swelling or bleeding in ~24% (lecanemab) and >33% (donanemab), adverse-event rates of 45% and 89%, and an NNH of about 3. It is the paper’s worked example of operation 2 applied to a fourth-row disease.

Design and delivery platforms

12. RFdiffusion — de novo design of protein structure and function; and atomically accurate de novo design of antibodies. Evidence that computational protein design can now produce binders and antibodies against specified targets, shortening the development pipeline that Directions 5 and 8 depend on.
13. In vivo CAR T-cell generation to treat cancer (and autoimmune disease) — targeted lipid nanoparticle delivery. The engineering step the paper identifies as the shared unlock for operations 3 and 4: generating engineered cells inside the patient, removing external per-patient manufacture, and collapsing the cost that currently blocks both rare-disease gene therapy and widespread cellular reset.

Atherosclerotic cardiovascular disease

14. CANTOS — canakinumab trial (summaries by WikiJournalClub and the American College of Cardiology). The 10,061-patient IL-1β trial in post-MI patients with hsCRP ≥ 2 mg/L. Source of the entry criteria, the per-dose hazard ratios (0.85 for 150 mg; 0.86 for 300 mg), the prespecified alpha thresholds and the 0.0004 significance margin, the LDL result (+1.7% with fewer events), and the harms (fatal infection/sepsis 0.31 vs 0.18 per 100 person-years). It grounds both the classification of cardiovascular disease in the fourth row and Direction 10’s warning about spending statistical space on doses instead of strata.
15. The CANTOS on-treatment hsCRP analysis (Residual Risk Reduction Initiative; TCTMD). Source of the stratified hazard ratios and the numbers needed to treat by hsCRP state (NNT 16 below 2 mg/L versus 57 above; all-cause mortality HR 0.69 versus 1.05). It shows a division strong enough to reverse the sign of mortality — while the paper itself flags that, being a post-treatment division, it is weaker evidence than HARP-2 or SAVE-MORE.

Oncology

16. FLAURA — first-line osimertinib in EGFR-mutated non-small-cell lung cancer (Cancer Management and Research; Journal of Oncology Navigation and Survivorship). Source of the intra-patient resistance argument: median progression-free survival 18.9 versus 10.2 months (HR 0.46), median overall survival 38.6 versus 31.8 months (HR 0.80), and 28% versus 9% still on regimen at three years — meaning 72% off it. The paper reads this as the AMR curve compressed into one body and eighteen months, placing cancer in the fourth row.

Framework

17. J. Konstapel, The Vacuum.Net Theory, Ninth Edition (Constable Research, Leiden, 27 October 2026) — Foundations 9.2 and Physics 9.0. The author’s underlying theoretical framework from which the four-operation scheme derives.
18. J. Konstapel, Where Illness Comes From, Completed (Constable Research, 20 October 2026). Source of the “cure statuses” (self-limiting; curable by removal; curable by supply; relief only) used to classify conditions into the four rows.
19. J. Konstapel, Depth of Discharge (Constable Research, 29 October 2026). The companion work that sets out the stratification argument in full — the basis for Directions 7 and 10 and for the claim that null fourth-row trials average away real effects.

Primary source

20. J. Konstapel, “Research Directions to Cure All Diseases,” Constable Research, Leiden, 29 October 2026. The paper this essay is based on. It sets out the four-operation classification, the evidence for each row, the ten research directions, and its own verification and limitations section.

J.Konstapel, Leiden,29-10-2026.

Article