Michael Quoc Founder & CEO · Axioms.ai
Vector 1 of 4: Causal hallmarks of aging

Axiomatic Intelligence · An investigation

The Physics of Longevity

No instrument on the body, and no biological-age test, reads the rate of aging. The body can still be read as a system with instruments, within limits: each instrument reads a defined quantity with measured noise and feeds a decision written down before the reading. Blood panels stand as targeted instruments read against each analyte's reference change value; a fixed quarterly cadence over a broad, unplanned panel has only weak evidence against it, so this page drops the word quarterly and does not yet cite a false-alarm count. Continuous glucose monitoring is a meaningful instrument for aggregate and threshold quantities over days under one sensor brand; ranking single meals by it is refuted, the quoted variability hazard ratios belong to years of laboratory fasting glucose, and eating to the trace has no outcome evidence in a non-diabetic adult. Sleep staging holds for duration, timing and regularity, and nightly deep and REM minutes are refuted. HRV holds as a rolling multi-day baseline under one device and protocol; a single day's reading, and any readiness or aging reading, are refuted. 'The question is never how do I feel, it is what does the data say' stays unresolved, with one bound: the sentence needs the clause which data. 'Most of longevity is not exotic: sleep, zone 2, strength, protein, no alcohol, no sugar' needs a re-verified source set before this page states it. Its direction survives as a statement about where attributable burden concentrates in cohort studies, but the word most has no defensible number; zone 2 resolves to aerobic volume at any intensity with no evidence of a unique benefit; strength is upheld for function and not demonstrated for mortality; protein is adequacy near a measured plateau in people who train; no sugar means no sweetened drinks, acting on liver fat; no alcohol is upheld in direction with the zero-versus-occasional difference inside the noise; sleep is unresolved because its healthspan evidence was not gathered in this round; and the list omits energy balance, blood pressure, smoking and glycemia, which outrank several of the named levers. 'The rest is measurement' is unresolved, bounded only indirectly: a change below an instrument's floor is not information.

The question this page asks: what actually moves human healthspan, and which of the measurements an engineer can take at home mean anything?

“I approach health the way I approach engineering - with data, protocols, and first-principles reasoning. Nutrition, hormone balance, sleep architecture, supplementation, metabolic markers. I want to understand the systems, not just follow the defaults.”

“The body is the platform everything else runs on, and I invest in it accordingly.”

This page rests on four research vectors and 32 axioms, forged on October 4, 2026 under the current version of the method. Every axiom lists the mechanism, the underlying physics, and the primary sources it rests on. Nothing on this page is taken on faith; it can be read one piece at a time.

Which proposed hallmarks of aging have human causal evidence, and what separates a causal aging mechanism from a disease-specific effect, a biomarker or a mouse result?

Causal hallmarks of aging · eight axioms

The mechanism under test is the hallmarks of aging: which proposed mechanisms are causally identified in humans and which are disease effects, biomarkers or animal results. The easy answer this vector was told not to settle for: “The hallmarks are established causes, so reversing one reverses aging”
1.1 A hallmark earns "cause of human aging" only through an aging-rate estimand; every human hallmark trial to date closed at a disease or a surrogateVERIFIEDLowering one age-related cause of death shifts that disease's survival curve without changing how fast total hazard doubles with age, so disease causation never implies aging-rate causation.Open this axiomClose this axiom
The axiom in full

Age-specific mortality in adult human populations rises roughly exponentially with age, so an intervention can move the hazard's level (delay or prevent a disease; a constant hazard ratio across age strata) or its doubling rate (slow aging), and the two are different quantities. The twelve-hallmark framework (S1) admits a mechanism on three criteria met largely in animal models and does not itself claim a human aging-rate effect; the independent audit of its evidence (S2) finds reliance on lifespan as a proxy, models of unclear relevance and designs without young-treated arms. In the human record this run opened, no perturbation of any hallmark has moved an aging-rate estimand: CANTOS moved a cardiovascular composite in 10,061 post-infarction patients and not all-cause mortality or incident frailty (S9, S10); tafamidis moved all-cause mortality in 441 patients with one protein-misfolding disease (S21); a senolytic pair moved one organ's histology in 31 patients with MASH (S15); CALERIE moved one methylation clock and not two others in 220 randomized adults (S7). Population: human adults in randomized trials and genetic cohorts. Outcome class: all-cause mortality and multi-system function or frailty, against single-disease incidence or a molecular surrogate. Evidence class: INTERVENTION and ARGUMENT. Boundary: this grades identification, not biology; a mechanism can be causal for a disease and still be the right target of a disease trial.

The physics underneath

A lever that lowers one cause of death leaves the other causes and the shared repair machinery untouched, so that disease's survival curve shifts (a level change) while the age-dependence of total hazard (the rate) is unchanged. Observed: the endpoints each trial moved and did not move. Inferred: the level-versus-rate decomposition as the estimand that separates them.

The sources
This is the first axiom on the page.Next axiom: 1.2, Blocking IL-1β in 10,061 post-infarction patients cut recurrent cardiovascular events about 15 percent and left all-cause mortality (HR 0.94) and incident frailty (HR 1.03) unchanged; circulating CRP is a readout, not the cause →Link to this axiom: #ax-1-1
1.2 Blocking IL-1β in 10,061 post-infarction patients cut recurrent cardiovascular events about 15 percent and left all-cause mortality (HR 0.94) and incident frailty (HR 1.03) unchanged; circulating CRP is a readout, not the causeVERIFIEDNeutralizing IL-1β lowers plaque inflammation and recurrent cardiovascular events but raises fatal infection through the same pathway, so all-cause death and frailty stay flat while CRP only reports the signal.Open this axiomClose this axiom
The axiom in full

In CANTOS (S9), 10,061 patients with a prior myocardial infarction and high-sensitivity CRP of 2 mg/L or more were randomized to canakinumab or placebo and followed a median 3.7 years; the 150 mg arm reduced the primary composite (myocardial infarction, stroke or cardiovascular death) about 15 percent (hazard ratio 0.85), hsCRP fell 26 to 41 percent dose-dependently, fatal infection or sepsis rose to 0.31 from 0.18 per 100 person-years (p=0.02), and all-cause mortality did not differ (hazard ratio 0.94, 95 percent CI 0.83 to 1.06). In the post hoc frailty analysis of 9,942 CANTOS participants with baseline data (S10), 12.5 percent of the 8,620 non-frail at baseline became frail on a 34-item cumulative-deficit index (score above 0.2) over 5 years, and allocation to canakinumab did not change that incidence (hazard ratio 1.03, 95 percent CI 0.91 to 1.17, p=0.63). In 194,418 participants across 47 studies with 46,557 coronary cases (S11), the coronary risk ratio per standard deviation of genetically raised log-CRP was 1.00 (95 percent CI 0.90 to 1.13) against an observational 1.33 (1.23 to 1.43). Population: adults with established atherosclerotic disease and elevated hsCRP; for the genetic result, general adult populations. Outcomes: MACE (major adverse cardiovascular events), all-cause mortality, incident frailty, CHD. Evidence class: INTERVENTION (S9, S10), CORRELATIONAL genetic instrument (S11). Boundary: holds for canakinumab in secondary prevention; it does not grade IL-6 blockade or diet and exercise effects on CRP, which are different exposures.

The physics underneath

IL-1β signaling drives plaque inflammation and thrombosis, so neutralizing it lowers atherothrombotic events; the same signaling is load-bearing for host defence, so fatal infections rise and the net death ledger does not move; the cumulative deficits that make up frailty are not downstream of this one cytokine. CRP rises with the pathway but has no causal weight of its own on coronary disease under the genetic instrument. Observed: the hazard ratios, infection rate, frailty incidence and CRP risk ratios. Inferred: that benefit and harm draw on one shared pathway.

The sources
← Previous axiom: 1.1, A hallmark earns "cause of human aging" only through an aging-rate estimand; every human hallmark trial to date closed at a disease or a surrogateNext axiom: 1.3, Under genetic instruments, longer telomeres raise cancer risk and lower some cardiovascular and neurodegenerative risks, and move neither lifespan nor late-life physical or cognitive function →Link to this axiom: #ax-1-2
1.3 Under genetic instruments, longer telomeres raise cancer risk and lower some cardiovascular and neurodegenerative risks, and move neither lifespan nor late-life physical or cognitive functionVERIFIEDLonger telomeres extend proliferative reserve, which protects degenerating tissue and feeds malignant clones alike, so disease risks shift in opposite directions and lifespan does not move.Open this axiomClose this axiom
The axiom in full

Across 35 cancers and 48 non-neoplastic diseases in 420,081 cases and 1,093,105 controls summed across outcomes (S3), genetically longer telomeres raised the odds of 9 of 22 primary cancers and lowered the odds of 6 of 32 primary non-neoplastic diseases (coronary heart disease odds ratio 0.78, 95 percent CI 0.67 to 0.90; abdominal aortic aneurysm 0.63, 0.49 to 0.81; Alzheimer disease 0.84, 0.71 to 0.98; each per standard deviation of genetically predicted telomere length). In 261,837 UK Biobank participants aged 60 and over followed a mean 7.5 years with a 13-variant instrument (S4), each 250 base pairs of genetically longer telomere (about a decade's average attrition) raised cancer odds (odds ratio 1.11, 95 percent CI 1.06 to 1.16), lowered coronary heart disease odds (0.95, 0.92 to 0.98), and showed little association with cognitive function, grip strength, sarcopenia or falls. With 153 variants in about 500,000 UK Biobank participants (S5), genetically longer telomere length was not associated with lifespan in men (0.17 years per standard deviation of log telomere length, 95 percent CI -0.54 to 0.88) or in women (0.04, -0.88 to 0.96). A systematic review of 62 Mendelian randomization studies covering 310 outcomes (S26) reports the same bidirectional pattern. Population: European-ancestry adults in biobanks and case-control consortia. Outcomes: disease odds, lifespan, late-life function. Evidence class: CORRELATIONAL (genetic instruments). Boundary: these are lifelong, small, developmentally compensated shifts in leukocyte telomere length; they do not identify the effect, safety or reversibility of lengthening telomeres in an older adult, and leukocyte length need not be the operative variable in every tissue.

The physics underneath

Telomere length sets proliferative reserve in replicating tissues; more reserve preserves vascular and regenerative tissue (fewer degenerative events) and permits more divisions in incipient malignant clones (more cancer), so the effects offset across causes and the net on lifespan is zero within the confidence bounds. Observed: the odds ratios and lifespan estimates. Inferred: antagonistic pleiotropy as the structure behind the sign flip.

The sources
← Previous axiom: 1.2, Blocking IL-1β in 10,061 post-infarction patients cut recurrent cardiovascular events about 15 percent and left all-cause mortality (HR 0.94) and incident frailty (HR 1.03) unchanged; circulating CRP is a readout, not the causeNext axiom: 1.4, DNA-methylation clocks are instrument outputs that respond discordantly to one randomized diet intervention and are not enriched for causal methylation sites; a clock's change is not a mediator's change →Link to this axiom: #ax-1-3
1.4 DNA-methylation clocks are instrument outputs that respond discordantly to one randomized diet intervention and are not enriched for causal methylation sites; a clock's change is not a mediator's changeVERIFIEDEach methylation clock is fit to a different target, so one real intervention moves the clocks in different directions, which shows the clocks register the perturbation rather than mediate aging.Open this axiomClose this axiom
The axiom in full

In CALERIE (S8), 218 adults without obesity were randomized to 25 percent caloric restriction or ad libitum intake for 2 years and the restriction arm achieved 11.7 percent (plus or minus 0.7) restriction of baseline intake; in the methylation analysis of 220 randomized participants (S7), the intervention slowed DunedinPACE (Cohen's d about 0.3 at 12 months and 0.2 at 24 months, p below 0.01, in the 197-participant conference analysis) and did not significantly change PhenoAge or GrimAge, with other clocks moving in the opposite direction. In an epigenome-wide Mendelian randomization of methylation sites against lifespan, healthspan and frailty (S6), neither the existing clocks nor age-related differential methylation were enriched for the putatively causal CpG sites. In 28 cohorts (n=34,710) with genetic instruments for GrimAge, PhenoAge, HannumAge and intrinsic epigenetic age acceleration tested against parental attained age and survival to recruitment (S5), all four clocks showed null associations with lifespan in both sexes. The authors' observational mapping of 3 percent slower DunedinPACE to 15 percent lower mortality in a different cohort is an association, not a trial effect (S25). In 951 Dutch Hunger Winter Families Study participants (S27), in-utero famine exposure associated with faster DunedinPACE six decades later: a clock registers a developmental exposure as a readout. Population: non-obese adults in a trial; biobank adults; a famine birth cohort. Outcomes: clock scores, lifespan. Evidence class: INTERVENTION (S7, S8), CORRELATIONAL genetic (S5, S6), CORRELATIONAL natural experiment (S27), ARGUMENT method (S25). Boundary: the clocks remain valid cohort predictors of mortality, which is their training; this bounds their use as a within-person outcome for an intervention and does not grade the causality-enriched clocks S6 proposes.

The physics underneath

Each clock is a regression fit to a different training target (chronological age, a mortality composite, a longitudinal pace), so each weights a different set of CpG sites; a perturbation moves the sites it reaches and leaves the others, and the clocks disagree in direction. A surrogate becomes a mediator only when the intervention's effect on the surrogate carries its effect on the outcome in the same trial (Prentice's criterion), which no clock has met. Observed: the clock effect sizes and signs, the non-enrichment, the null lifespan MR, the famine association. Inferred: that the clocks measure the perturbation rather than the construct.

The sources
← Previous axiom: 1.3, Under genetic instruments, longer telomeres raise cancer risk and lower some cardiovascular and neurodegenerative risks, and move neither lifespan nor late-life physical or cognitive functionNext axiom: 1.5, Senolytics engage their target in human tissue within days, and no prespecified primary clinical endpoint has moved except in one organ-specific trial of 31 patients; the mouse result that admitted the hallmark was a 24 to 27 percent median-lifespan gain →Link to this axiom: #ax-1-4
1.5 Senolytics engage their target in human tissue within days, and no prespecified primary clinical endpoint has moved except in one organ-specific trial of 31 patients; the mouse result that admitted the hallmark was a 24 to 27 percent median-lifespan gainVERIFIEDEliminating senescent cells relieves fibrotic or resorptive signaling only where their burden is pathologically high, so human benefit appears in one diseased organ, not in healthy tissue.Open this axiomClose this axiom
The axiom in full

In 9 participants with diabetic kidney disease given a 3-day course of the dasatinib-plus-quercetin pair, open-label with no control (S12), adipose p16-positive cells fell 35 percent, p21-positive cells 17 percent and senescence-associated beta-galactosidase activity 62 percent, with circulating IL-1α, IL-6, MMP-9 and MMP-12 lower at 11 days: target engagement in human tissue. In 14 patients with idiopathic pulmonary fibrosis, single-arm and open-label over 3 weeks (S13), 6-minute walk distance rose 21.5 meters and gait speed and chair stands improved, while FEV1, FVC, the frailty index and SASP markers did not change. In 60 postmenopausal women randomized to intermittent treatment or control for 20 weeks (S14), the primary endpoint, percent change in the bone-resorption marker CTx, did not differ (-4.1 percent against -7.7 percent, p=0.611); only the top tertile of baseline T-cell p16 showed exploratory responses. In 31 patients with fibrotic MASH randomized double-blind to intermittent treatment or placebo across three 7-week cycles (S15), 47 percent against 7 percent reached at least one stage of fibrosis improvement without MASH worsening on paired biopsies (p=0.02), a one-organ histological endpoint over 21 weeks. The mouse result that admitted the hallmark (S16): eliminating p16-positive cells from 12 months of age in INK-ATTAC mice raised median lifespan 24 to 27 percent in a mixed background and 24 percent in C57BL/6. Population: patients with one named disease (kidney, lung, bone, liver); mice for the founding result. Outcomes: tissue senescence markers, organ-specific function and histology, a bone-turnover surrogate. Evidence class: INTERVENTION (open-label S12, S13; randomized S14, S15; mouse S16). Boundary: human evidence holds only for disease-defined cohorts with high senescent burden; no trial has a young-treated arm, a multi-system endpoint or an all-cause ledger.

The physics underneath

Senescent cells accumulate where injury and replication stress concentrate and secrete a pro-inflammatory, matrix-remodeling program; eliminating them relieves that organ's fibrosis or resorption signal where the burden is high and yields nothing measurable where the burden is ordinary; a multi-kinase inhibitor plus a flavonoid cannot attribute the effect to senescent-cell loss without a rescue arm or a mediator dose-response. Observed: the marker reductions, the null CTx primary, the fibrosis response rates, the mouse lifespan gain. Inferred: the threshold dependence and the attribution problem.

The sources
← Previous axiom: 1.4, DNA-methylation clocks are instrument outputs that respond discordantly to one randomized diet intervention and are not enriched for causal methylation sites; a clock's change is not a mediator's changeNext axiom: 1.6, Human perturbations of nutrient sensing move immune surrogates and disease incidence without moving lifespan: mTOR inhibition and lifelong growth-hormone-receptor loss both prove the pathway reachable and neither closes the aging chain →Link to this axiom: #ax-1-5
1.6 Human perturbations of nutrient sensing move immune surrogates and disease incidence without moving lifespan: mTOR inhibition and lifelong growth-hormone-receptor loss both prove the pathway reachable and neither closes the aging chainVERIFIEDDamping growth-hormone, IGF-1 and mTOR signaling lowers proliferation-driven disease and restores an immune surrogate, but the same signaling supports repair, so lifespan and clinical primaries do not move.Open this axiomClose this axiom
The axiom in full

In 218 older adults randomized to low-dose everolimus or placebo (S18), the influenza vaccine response rose about 20 percent and the share of PD-1-positive CD4 and CD8 T cells fell: an immune surrogate moved. In the phase 3 trial of the mTOR inhibitor RTB101 in 1,024 older adults over a 16-week winter season (S19), clinically symptomatic respiratory illness occurred in 26 percent on drug against 25 percent on placebo (p=0.65) despite upregulated interferon-induced antiviral genes, after a phase 2b of 652 had shown 19 percent against 28 percent laboratory-confirmed infections (p=0.02). In PEARL (S20), 114 completers aged 50 to 85 randomized to weekly compounded rapamycin at two doses or placebo for 48 weeks showed no change in the primary endpoint, visceral adiposity by DXA (partial eta squared 0.001, p=0.942); lean mass and self-reported pain improved only among women on the higher dose, a subgroup of 8 on secondary endpoints. In the Ecuadorian growth-hormone-receptor-deficiency cohort (S17), 99 individuals with lifelong IGF-1 deficiency against 1,606 unaffected relatives had one cancer death against cancer as 20 percent of relatives' deaths, and 0 percent diabetes (95 percent CI 0 to 4) against about 5 to 6 percent; serum from affected individuals lowered oxidant-induced DNA damage in cultured cells, an effect fully reversed by adding 200 ng/mL IGF-1 back, a human ex vivo add-back that identifies the mediator. No lifespan extension was demonstrated in that cohort and childhood mortality was elevated. Population: healthy older adults in trials; a lifelong loss-of-function kindred. Outcomes: vaccine titers, respiratory infection incidence, visceral fat, cancer and diabetes incidence, lifespan. Evidence class: INTERVENTION (S18, S19, S20), CORRELATIONAL natural cohort with an ex vivo INTERVENTION (S17). Boundary: trial results hold for the doses, durations and populations named; compounded rapamycin in PEARL had about one third the bioavailability of the commercial form, so a null there is a null for that exposure; the add-back identifies IGF-1 as the mediator of the cell phenotype, not of lifespan.

The physics underneath

Growth-hormone, IGF-1 and mTOR signaling scale anabolism and proliferation; damping them lowers the diseases of proliferation and glucose load (cancer, diabetes) and partially restores an exhausted immune repertoire, while the same signaling supports growth, glucose handling and repair, so a complete lifelong knockdown takes away two causes of death without extending life and a pharmacological knockdown moves a surrogate without moving a clinical primary. Observed: the titers, infection rates, the null primary, the cancer and diabetes counts, the reversal by IGF-1. Inferred: that disease causation here does not entail aging-rate causation.

The sources
← Previous axiom: 1.5, Senolytics engage their target in human tissue within days, and no prespecified primary clinical endpoint has moved except in one organ-specific trial of 31 patients; the mouse result that admitted the hallmark was a 24 to 27 percent median-lifespan gainNext axiom: 1.7, A genetic instrument identifies the effect of a lifelong, small, developmentally compensated shift under three assumptions and licenses a sign and a target's non-futility, never a dose, duration or age window; a somatic natural experiment trades that exogeneity for mechanistic relevance →Link to this axiom: #ax-1-6
1.7 A genetic instrument identifies the effect of a lifelong, small, developmentally compensated shift under three assumptions and licenses a sign and a target's non-futility, never a dose, duration or age window; a somatic natural experiment trades that exogeneity for mechanistic relevanceVERIFIEDVariants fixed at conception identify a lifelong, buffered exposure shift free of confounding but blind to dose and timing; somatic clones supply dose and timing but lose exogeneity.Open this axiomClose this axiom
The axiom in full

Mendelian randomization identifies an exposure's effect only when the variants predict the exposure (relevance), are unrelated to confounders (independence) and reach the outcome only through the exposure (exclusion); what it identifies is a lifelong small shift set at conception and buffered by development, so it licenses the sign of an effect and whether a target is futile, never the effect of a late-life, large, uncompensated change. The telomere record shows the dependence on instrument quality: a 13-variant instrument in 261,837 older adults (S4), a multi-trait set across 420,081 cases (S3) and a 153-variant set in about 500,000 participants (S5) agree, while the anthropic leg records a 7-variant score that once gave the opposite cardiovascular sign. The CRP result (S11) shows exclusion doing its work: the observational risk ratio 1.33 fell to 1.00 under the instrument. Somatic natural experiments reverse the trade: in four case-control cohorts (4,726 coronary cases, 3,529 controls) plus prospective cohorts (S22), carriers of clonal hematopoiesis had about 1.9-fold coronary heart disease risk (95 percent CI 1.4 to 2.7) and 4.0-fold early-onset myocardial infarction (2.4 to 6.7), with coronary calcification rising with the mutations and Tet2-deficient marrow enlarging aortic lesions in mice; in 97,691 whole genomes (S23) an inherited IL6R loss-of-function variant attenuated the cardiovascular risk associated with CHIP, a natural rescue arm for the proposed mediator. A somatic clone grows in the right tissue at the right age without developmental buffering, but clone size is endogenous to the inflammatory and proliferative stresses that also cause the outcome. Population: European-ancestry biobank adults; coronary case-control cohorts. Outcomes: disease odds, lifespan, coronary events. Evidence class: CORRELATIONAL (genetic and somatic instruments; S22 adds a mouse INTERVENTION). Boundary: neither design identifies a treatment effect; a drug's dose, duration and age window need a randomized trial; the IL6R attenuation estimate is not carried because only the paper's identity was confirmed.

The physics underneath

Randomization at conception removes confounding but fixes the exposure's size and timing to a lifelong, buffered shift; a somatic clone supplies the right dose, tissue and age window but is itself a product of the exposure field, so the two designs carry non-overlapping biases and their agreement, not either alone, is the evidence. Observed: the instrument-set estimates, the CRP contrast, the CHIP risk ratios and the IL6R attenuation. Inferred: the orthogonal-bias reading of germline against somatic designs.

The sources
← Previous axiom: 1.6, Human perturbations of nutrient sensing move immune surrogates and disease incidence without moving lifespan: mTOR inhibition and lifelong growth-hormone-receptor loss both prove the pathway reachable and neither closes the aging chainNext axiom: 1.8, A null treatment trial is not a null mechanism and a positive disease trial is not an aging result: identification attaches to the tuple of perturbation, dose, window, population and endpoint →Link to this axiom: #ax-1-7
1.8 A null treatment trial is not a null mechanism and a positive disease trial is not an aging result: identification attaches to the tuple of perturbation, dose, window, population and endpointVERIFIEDEach trial fixes one perturbation, dose, window, population and endpoint, so its positive or null result transfers only along the dimensions held fixed and never grades the hallmark label itself.Open this axiomClose this axiom
The axiom in full

In ATTR-ACT (S21), 441 patients with transthyretin amyloid cardiomyopathy randomized to tafamidis or placebo for 30 months had all-cause mortality of 29.5 percent against 42.9 percent (hazard ratio 0.70, 95 percent CI 0.51 to 0.96) and cardiovascular hospitalization 0.48 against 0.70 per year (rate ratio 0.68, 0.56 to 0.81): a proteostasis mechanism closes to survival in one disease, which is a disease result and not a slowing of aging. In MMPOWER-3 (S24), 218 people with genetically confirmed primary mitochondrial myopathy randomized to elamipretide or placebo for 24 weeks did not meet the primary walking and fatigue endpoints: a null for that peptide at that dose in that population, not evidence that mitochondrial defects are non-causal, because a trial can fail at the target, the tissue, the downstream function, the timing relative to irreversible damage, the balance of harms or the endpoint. Randomizing a treatment does not randomize the mediator; a post-treatment marker adjusted into an outcome regression does not identify mediation without further assumptions (S28), and a surrogate carries a treatment effect only when the treatment's effect on the surrogate accounts for its effect on the outcome in the same trial (S25). Population: patients with one named disease. Outcomes: all-cause mortality and hospitalization; functional endpoints. Evidence class: INTERVENTION (S21, S24), ARGUMENT method (S25, S28). Boundary: grades the logic of transfer; it does not grade tafamidis or elamipretide as treatments.

The physics underneath

A hallmark label names a family of variables and of possible interventions; each trial tests one member at one dose in one population against one endpoint, so its result transfers only along the dimensions it held fixed, and a positive or null reading of the label as a whole is a category error. Observed: the ATTR-ACT and MMPOWER-3 endpoints. Inferred: the failure-mode list for a null and the tuple rule, from the method sources.

The sources
← Previous axiom: 1.7, A genetic instrument identifies the effect of a lifelong, small, developmentally compensated shift under three assumptions and licenses a sign and a target's non-futility, never a dose, duration or age window; a somatic natural experiment trades that exogeneity for mechanistic relevanceNext axiom: 2.1, The reference change value is the floor on detectable serial change, and the database value is the optimistic edge of that floor →Link to this axiom: #ax-1-8

What do blood panels, biological-age clocks, continuous glucose monitors, heart-rate variability and consumer sleep staging estimate, with what error, and when does repeated testing improve a decision?

Instrument validity and utility · eight axioms

The mechanism under test is what an instrument measures: blood panels, biological-age clocks, continuous glucose monitors, HRV and consumer sleep staging as instruments with estimand, noise, cadence and decision utility (an audit track). The easy answer this vector was told not to settle for: “Quarterly bloodwork and more precise tracking necessarily improve longevity”
2.1 The reference change value is the floor on detectable serial change, and the database value is the optimistic edge of that floorVERIFIEDTwo draws differ by biological plus analytical noise, so a serial change below about 2.8 combined CVs is indistinguishable from no change, and routine collection widens that floor.Open this axiomClose this axiom
The axiom in full

For one person, two laboratory results differ by biological and analytical noise before any real change enters; the smallest difference distinguishable from no change at a 95 percent criterion is RCV = 1.96 x sqrt(2) x sqrt(CV_A^2 + CV_I^2), about 2.8 combined single-result CVs. Population: healthy adults sampled under protocol; outcome: a biomarker's change against its test-retest noise; design: within-person biological-variation studies (CORRELATIONAL). Measured components: creatinine within-subject CV 5.2 percent (95 percent CI 4.6 to 5.8) pooled from 37 studies and 2,770 participants, giving an eGFR RCV of plus or minus 12.5 to 16.5 percent (S3, 2022); LDL cholesterol within-subject CV 8.7 percent by direct assay, 9.3 by Friedewald, 9.0 by Martin-Hopkins in healthy adults, which with an analytical CV near 2.5 percent gives a two-result RCV near 26 percent of the value (S5, 2026; the 26 percent is arithmetic from the published components); HbA1c within-individual total variation 0.091 in non-diabetic primary-care patients among 587,023 patients, over five times the EFLM database biological-variation figure of 0.016, because routine collection adds pre-analytical variation that controlled studies exclude (S6, 2025); indirect RCVs from routine laboratory data are similar to but wider than formula RCVs, creatinine minus 16 to plus 14 percent (S8, 2024). Of 61 biological-variation papers appraised for kidney markers, 3 earned grade A, 4 B and 48 C (S4), so database inputs carry their own uncertainty. Boundary: the protocol CV_I is a lower bound on the variation a person meets in routine collection; a population reference interval (CV_G) is not an individual's change threshold.

The physics underneath

The variance of a difference between two independent draws is the sum of the two variances (the sqrt 2); each draw's variance is the sum of independent biological and analytical components (the root-sum-square); a reference interval describes between-person spread, not one person's set point. Observation: the CV components and the real-world HbA1c figure. Inference: the RCV arithmetic and the lower-bound claim.

The sources
← Previous axiom: 1.8, A null treatment trial is not a null mechanism and a positive disease trial is not an aging result: identification attaches to the tuple of perturbation, dose, window, population and endpointNext axiom: 2.2, A broad panel at a fixed cadence without prespecified analytes manufactures false alarms by multiplicity and apparent cures by regression to the mean →Link to this axiom: #ax-2-1
2.2 A broad panel at a fixed cadence without prespecified analytes manufactures false alarms by multiplicity and apparent cures by regression to the meanVERIFIEDMany tests at a fixed cadence compound 5 percent false flags multiplicatively, and a flagged extreme reverts toward the mean on retest, so unprespecified panels manufacture both alarms and cures.Open this axiomClose this axiom
The axiom in full

In healthy adults, each analyte judged against a 95 percent reference interval flags about 5 percent of the time with no disease present, so k analytes read t times a year flag at least once with probability 1 minus 0.95^(kt) under independence: 20 analytes once, 64 percent; 20 analytes quarterly, 98 percent and four expected flags a year (MODEL arithmetic; correlation among analytes lowers the real figure but not to a small one). The measured analogue in repeated screening is the PLCO trial: among 68,436 adults aged 55 to 74, the cumulative risk of at least one false-positive result after 14 screening tests was 60.4 percent in men and 48.8 percent in women, and 28.5 percent of men and 22.1 percent of women underwent an invasive procedure prompted by a false positive (S9, 2009; CORRELATIONAL within a randomized trial; outcome: false-positive and procedure incidence). The second engine is regression to the mean: with between-person reliability rho = CV_G^2 / (CV_G^2 + CV_I^2 + CV_A^2), a value one SD above the mean reverts a fraction 1 minus rho on retest with no intervention; from LDL CV_G 18.6 and CV_I 9.0 percent rho is about 0.80 (S5), and from HbA1c CV_G 7.9 percent (S35) against the real-world non-diabetic within-person 9.1 percent (S6) rho is about 0.43, so more than half of an apparent HbA1c elevation disappears on retest by arithmetic (the rho values are this writer's computation from published components). The only quarterly rationale in an opened guideline is disease management: HbA1c at least quarterly when therapy changes or goals are unmet, twice yearly when stable (S34, 2023).

The physics underneath

Independent tests compound their false-flag probabilities multiplicatively, and a selected extreme value carries a noise component that is not expected to repeat. Observation: the PLCO counts and the CV components. Inference: the binomial and rho arithmetic, and the reading of the guideline cadence as disease management rather than surveillance.

The sources
← Previous axiom: 2.1, The reference change value is the floor on detectable serial change, and the database value is the optimistic edge of that floorNext axiom: 2.3, Within-person variation grows with the sampling interval for drifting analytes, so replicates clustered at decision points detect change better than singletons spread across the calendar →Link to this axiom: #ax-2-2
2.3 Within-person variation grows with the sampling interval for drifting analytes, so replicates clustered at decision points detect change better than singletons spread across the calendarVERIFIEDDrift enters within-person variance as the interval lengthens, so spaced singletons grow noisier while averaged replicates at decision points shrink noise by one over root n.Open this axiomClose this axiom
The axiom in full

In healthy adults, the within-person CV of a drifting analyte measured monthly exceeds its CV measured daily: for small dense LDL cholesterol the within-subject CV was 9.4 percent on daily fasting draws and 17.1 percent on monthly draws, with monthly within-subject CVs of 17.1 percent (sdLDL-C), 7.4 (HDL3-C) and 25.7 (TRL-C) and a monthly RCV of 18.1 mg/dL against a mean near 28 mg/dL (S7, 2022; CORRELATIONAL within-person; outcome: biomarker change against its noise at two intervals). Because the RCV of a mean of n replicates falls as 1 over sqrt n while CV_I rises with the interval, two draws a week apart averaged cut LDL's single-pair RCV from about 26 to about 18 percent and three to about 15 percent (arithmetic from the S5 components), whereas four quarterly singletons each carry the full interval-inflated CV_I. The analyte's integration time sets the shortest useful interval: HbA1c integrates glycation over erythrocyte survival, which is why the opened guideline's shortest cadence is about three months (S34). Boundary: shown directly for lipid subfractions in healthy adults; the generalization to other drifting analytes is inference from the variance model.

The physics underneath

Slow drift and seasonal components enter the within-person variance only when the interval is long enough for them to act; averaging independent replicates shrinks random error but not drift. Observation: the daily-versus-monthly CVs. Inference: the clustered-replicate design rule.

The sources
← Previous axiom: 2.2, A broad panel at a fixed cadence without prespecified analytes manufactures false alarms by multiplicity and apparent cures by regression to the meanNext axiom: 2.4, Epigenetic clocks are technically reliable after principal-component reconstruction, biologically unstable over days, prognostic at the cohort level and unproven as surrogates; the individual-tracking use is refuted on current evidence →Link to this axiom: #ax-2-3
2.4 Epigenetic clocks are technically reliable after principal-component reconstruction, biologically unstable over days, prognostic at the cohort level and unproven as surrogates; the individual-tracking use is refuted on current evidenceVERIFIEDPrincipal components remove array noise from methylation clocks but day-to-day biology still moves the inputs, so within-person change below about a quarter pace-unit is noise.Open this axiomClose this axiom
The axiom in full

Technical noise causes six major clocks to deviate 3 to 9 years between replicate assays of one sample, and principal-component versions bring most replicates within 1 year (S10, 2021; S11, 2022; MODEL on technical replicates); DunedinPACE test-retest ICC is 0.96 (95 percent CI 0.93 to 0.98) in 36 replicates and 0.87 in 350 (S13, 2022). Across 18 methylation clocks in four datasets (technical replicates in 196 and 128 individuals; biological replicates under meals, stress, altitude and pollution in 21 to 34 per condition), most clocks had technical ICC above 0.90 but biological short-interval ICC of about 0.4 to 0.7 with none excellent, and the two reliabilities were independent (r 0.0168) (S12, Aging Cell, July 2026; CORRELATIONAL). Prognosis: among 2,105 NHANES adults aged 50 and over followed a median 17.5 years with 998 deaths, GrimAge acceleration carried a mortality hazard ratio of 1.50 (1.32 to 1.71) per 5 years and Hannum, PhenoAge and Horvath 1.13 to 1.16 (S15, 2024 preprint; outcome: all-cause mortality). Surrogacy: in CALERIE, 220 adults without obesity randomized to 25 percent calorie restriction or ad libitum for two years, DunedinPACE slowed while PhenoAge and GrimAge did not change significantly, with small effects (S14, 2023; INTERVENTION); no biomarker of aging is a validated surrogate endpoint (S16, 2024; S2). Smallest real change in one person (MODEL arithmetic): MDC95 = 2.77 x SD x sqrt(1 minus ICC); with a DunedinPACE SD of 0.15 pace-units and a biological ICC of 0.70, MDC95 is about 0.23 pace-units, two to three times the 10 percent unit in which its mortality association is expressed. Boundary: cohort prognosis holds for middle-aged and older adults; the individual refutation holds for repeat tests over months to a year in any adult.

The physics underneath

A clock is a regression over methylation at selected CpG sites in a leukocyte mixture; averaging across principal components removes array noise but not the day-to-day biology (cell composition, meals, stress) that moves its inputs; the attenuation theorem caps a measure's correlation with any truth at the square root of its ICC. Observation: the ICCs, hazard ratios and CALERIE effects. Inference: the MDC arithmetic and the refutation of the individual-feedback use.

The sources
← Previous axiom: 2.3, Within-person variation grows with the sampling interval for drifting analytes, so replicates clustered at decision points detect change better than singletons spread across the calendarNext axiom: 2.5, A continuous glucose monitor estimates encapsulated interstitial glucose with a lag whose dominant part is the sensor's foreign-body response, and its accuracy in non-diabetic wearers is worse than the pivotal diabetes figure →Link to this axiom: #ax-2-4
2.5 A continuous glucose monitor estimates encapsulated interstitial glucose with a lag whose dominant part is the sensor's foreign-body response, and its accuracy in non-diabetic wearers is worse than the pivotal diabetes figureVERIFIEDThe sensor reads glucose through a growing collagen capsule, so interstitial readings lag and smear blood spikes, and non-diabetic wearers get one reading in three outside 15 percent.Open this axiomClose this axiom
The axiom in full

The physiological delay of glucose from blood to interstitial fluid in overnight-fasted adults is about 5 to 6 minutes by tracer (S17, 2013; INTERVENTION in humans), but in vivo sensors reach maximum current 24 plus or minus 7 minutes after a glucose rise against 2 plus or minus 1 minutes for blood and 35 plus or minus 2 seconds in vitro, and the delay correlates with collagen deposited around the filament (R 0.96), so encapsulation, which grows over the wear period, is the dominant lag (S18, 2019; INTERVENTION in mice; transfer to humans is inference). MARD is a mean absolute relative difference, not an error bound; the error bound is the agreement rate. In the pivotal study of 316 adults with diabetes, MARD was 8.2 percent (arm) and 9.1 percent (abdomen), with 89.6 and 85.5 percent of readings within 15 mg/dL or 15 percent of reference and 95.3 and 93.2 percent within 20/20 (S19, 2022; controlled clinic comparison), so about one reading in ten misses by more than 15. In 39 young adults without diabetes wearing over-the-counter sensors against a capillary meter, MARD was 12.9 percent, bias minus 4.8 mg/dL, 66.5 percent of readings within 15/15, population lag 6.2 minutes, and MARD rose from 10.6 to 15.3 percent after sugar versus placebo (S20, September 2026; the reference was a meter, not laboratory plasma). Boundary: the pivotal figures describe adults with diabetes in clinic; the 39-person figures describe young non-diabetic adults against a meter.

The physics underneath

The sensor reads current at an electrode inside a fibrous capsule in subcutis, three substitutions from plasma glucose (compartment, filter, encapsulation); a sensor with time constant tau cannot resolve a transient faster than tau plus its filter delay, and accuracy computed over stationary data understates error during the excursions a wearer cares about. Observation: the lag, MARD and agreement figures. Inference: that peak height, timing and width on a trace are systematically distorted and that the distortion changes across wear.

The sources
← Previous axiom: 2.4, Epigenetic clocks are technically reliable after principal-component reconstruction, biologically unstable over days, prognostic at the cohort level and unproven as surrogates; the individual-tracking use is refuted on current evidenceNext axiom: 2.6, A food's glucose effect on one person is not identifiable from one sensor and one exposure, because within-person meal response varies about as much as its mean and the device is part of the estimand; aggregate two-week metrics carry their own instrument uncertainty →Link to this axiom: #ax-2-5
2.6 A food's glucose effect on one person is not identifiable from one sensor and one exposure, because within-person meal response varies about as much as its mean and the device is part of the estimand; aggregate two-week metrics carry their own instrument uncertaintyVERIFIEDOne person's meal response varies about as much as its mean and two sensors rank meals differently, so one exposure on one sensor identifies no food effect.Open this axiomClose this axiom
The axiom in full

In 16 adults without diabetes on an inpatient unit for 28 days, 27,489 simultaneous readings from two sensors of different brands gave within-subject coefficients of variation of incremental postprandial glucose of 91.7 plus or minus 1.9 and 94.2 plus or minus 2.7 percent, and the two devices ranked the same meals with a mean Kendall tau of 0.43 (S21, 2020; within-person controlled feeding). In 394 healthy adults wearing two sensors for 14 days with 4,457 standardized and 5,738 free meals, two sensors of the same brand agreed on meal ranking at tau 0.87 with incremental-area CV 3.7 percent, while sensors of different brands agreed at tau 0.68 with CV 12.5 percent and misclassified 19 percent of meals between the top and bottom response quintiles (S22, 2022; within-person field design). In 172 adults with type 2 diabetes not on insulin wearing two sensors for two weeks at baseline and three months later, mean glucose between sensors correlated r 0.86 with RMSE 13 mg/dL and time-in-range r 0.88 with RMSE 8 percentage points, and within-person CV of mean glucose across the two periods was 17.4 and 14.2 percent by brand and of time-in-range 20.1 and 18.6 percent (S23, 2023; CORRELATIONAL). Power (MODEL arithmetic from S21): with within-subject CV near 92 percent, resolving a 25 percent difference between two foods in one person needs on the order of 150 to 210 paired exposures per food; a difference near one within-person SD, roughly a doubling, needs about 15 repeats. The only randomized outcome evidence for monitor-guided management is in disease: 175 adults with type 2 diabetes on basal insulin, adjusted HbA1c difference minus 0.4 points (95 percent CI minus 0.8 to minus 0.1) at eight months (S24, 2021; INTERVENTION); none exists in non-diabetic adults. Boundary: the meal-ranking result is from non-diabetic adults; the period-variability result is from type 2 diabetes not on insulin.

The physics underneath

The replication unit for a food question is the meal occasion, not the timestamp; dense adjacent readings are autocorrelated and count as far fewer independent observations than their number; when two instruments disagree on ranking, a share of any ranking is the instrument. Observation: the CVs, correlations and the trial effect. Inference: the exposure counts and the non-identifiability claim.

The sources
← Previous axiom: 2.5, A continuous glucose monitor estimates encapsulated interstitial glucose with a lag whose dominant part is the sensor's foreign-body response, and its accuracy in non-diabetic wearers is worse than the pivotal diabetes figureNext axiom: 2.7, RMSSD under a fixed protocol is a trend statistic with a within-person CV near 30 percent, so a single day's reading has no decision value and a multi-day baseline is the instrument; its demonstrated utility is training allocation, not outcome prediction →Link to this axiom: #ax-2-6
2.7 RMSSD under a fixed protocol is a trend statistic with a within-person CV near 30 percent, so a single day's reading has no decision value and a multi-day baseline is the instrument; its demonstrated utility is training allocation, not outcome predictionVERIFIEDBeat-interval differencing amplifies timing noise and daily state moves RMSSD about 30 percent, so only a multi-day baseline under one protocol carries a readable signal.Open this axiomClose this axiom
The axiom in full

In 50 healthy adults recorded 15 minutes at rest one week apart, heart rate had ICC 0.81 and CV 5.8 percent while RMSSD had ICC 0.75 and CV 29.6 percent (S25, 2025; test-retest). In 41 healthy adults over 14 days of standardized morning chest-strap recordings (424 observations), the within-person CV of RMSSD averaged 0.37 (SD 0.16, range 0.14 to 0.71), and higher RMSSD associated with better self-reported sleep (beta 0.51), lower fatigue (0.28) and lower stress (0.35) with modest effect sizes (S26, 2025; CORRELATIONAL). From a raw-scale CV of 30 percent, MDC95 = 1.96 x sqrt(2) x 30 percent is about 83 percent for a single day and about 31 percent for a seven-day mean under independence, worse with autocorrelation (MODEL arithmetic). Consumer nocturnal HRV against a chest-strap ECG in 13 healthy adults over 536 nights showed limits of agreement of minus 11.4 to plus 6.4 ms for one ring and minus 14.3 to plus 23.6 ms for one watch (S27, 2025; device validation; repeated nights sharpen the estimate for 13 people, not a population). Utility: in a randomized trial of 55 recreationally active adults, HRV-guided training (one-minute supine morning lnRMSSD against a 14-day baseline) produced the same fitness, body-composition and strength gains as predetermined training while completing 12.9 plus or minus 5.6 versus 26.5 plus or minus 2.6 high-intensity days (S28, 2021; INTERVENTION). Boundary: healthy and recreationally active adults; time-domain RMSSD under a fixed posture, window and time of day.

The physics underneath

RMSSD differences beat-interval timestamps, so independent timing error of variance sigma squared enters the successive-difference statistic as six sigma squared; respiration rate, posture, time of day, alcohol, illness and recent exertion are inputs to the quantity rather than confounders to adjust away, so the protocol is the measurement. Observation: the ICCs, CVs, limits of agreement and trial counts. Inference: the MDC arithmetic and the many-to-one claim that an RMSSD fall does not identify its cause.

The sources
← Previous axiom: 2.6, A food's glucose effect on one person is not identifiable from one sensor and one exposure, because within-person meal response varies about as much as its mean and the device is part of the estimand; aggregate two-week metrics carry their own instrument uncertaintyNext axiom: 2.8, Consumer sleep staging detects sleep and misses wake, classifies stages at fair-to-moderate agreement, and averaging nights sharpens a biased stage estimate rather than correcting it; duration and timing are upheld, nightly deep and REM minutes are refuted →Link to this axiom: #ax-2-7
2.8 Consumer sleep staging detects sleep and misses wake, classifies stages at fair-to-moderate agreement, and averaging nights sharpens a biased stage estimate rather than correcting it; duration and timing are upheld, nightly deep and REM minutes are refutedVERIFIEDWrist and bed sensors infer stages from motion and pulse without EEG, so quiet wake reads as sleep and stage errors are systematic, which averaging sharpens rather than removes.Open this axiomClose this axiom
The axiom in full

Against polysomnography, consumer devices call sleep well and wake poorly: in 34 healthy young adults over three laboratory nights with seven devices, sleep sensitivity was at least 0.93 on every device while wake specificity ranged 0.18 to 0.54, all devices overestimated light sleep and half misclassified deep or REM significantly (S29, 2021); in six wrist wearables, sleep sensitivity exceeded 90 percent with specificity 29.4 to 52.2 percent and kappa 0.21 to 0.53 (S33, 2018); in 98 adults, one watch had sensitivity 95.3 to 95.9 percent, specificity 44.5 to 47.9 percent, kappa 0.43 and a prevalence-and-bias-adjusted kappa of 0.75, which describes a hypothetical balanced night no one sleeps (S31, 2025). Stage agreement: across 11 trackers in 75 adults and 349,114 epochs, four-stage macro-F1 ranged 0.26 to 0.69, with kappa fair (0.2 to 0.4) for two leading wearables and moderate (0.4 to 0.6) for three others (S30, 2023). Pooled across 36 studies of contactless devices, total sleep time was overestimated by 19.55 minutes, deep sleep by 11.07 minutes and wake after onset underestimated by 12.07 minutes (S32, 2023; meta-analysis). The averaging law (MODEL): with true stage proportions pi and the device's confusion matrix C, expected reported proportions are C pi, so more nights make the mean more precise around a biased estimate, and an algorithm update or a change in sleep conditions invalidates any correction. Boundary: healthy and mixed clinical adults in laboratory nights; the asymmetry is worst where wake is the quantity of interest (insomnia, OSA).

The physics underneath

The device observes motion and pulse features and infers a 30-second epoch label with no electroencephalogram, so a stage is a classifier output whose errors are systematic (quiet wake looks like sleep; autonomic arousal after alcohol or late exercise looks like absent deep or REM), and systematic error does not average away. Observation: sensitivities, specificities, kappas and minute biases. Inference: the C pi law and the note that observing can perturb the state observed (orthosomnia).

The sources
← Previous axiom: 2.7, RMSSD under a fixed protocol is a trend statistic with a within-person CV near 30 percent, so a single day's reading has no decision value and a multi-day baseline is the instrument; its demonstrated utility is training allocation, not outcome predictionNext axiom: 3.1, Fitness predicts death; training moves fitness; the two coefficients are not the same number →Link to this axiom: #ax-2-8

How do exercise dose, intensity and modality relate to mortality and healthspan, and which comparisons with diet, sleep and pharmacology are defensible on matched endpoints?

Exercise dose response and comparison · eight axioms

The mechanism under test is exercise dose-response and the comparison across levers: fitness, strength and activity against mortality and healthspan, with the not-exotic claim's operational comparator. The easy answer this vector was told not to settle for: “Exercise is the strongest lever and zone 2 is its proven secret”
3.1 Fitness predicts death; training moves fitness; the two coefficients are not the same numberVERIFIEDFitness integrates genotype, disease and body composition as well as training, so its mortality gradient over-states what a training-induced fitness gain can move.Open this axiomClose this axiom
The axiom in full

In general-population cohorts the baseline cardiorespiratory fitness gradient is the largest mortality association in exercise epidemiology: high versus low CRF carries an all-cause mortality HR of 0.47 (95% CI 0.39 to 0.56) and each 1-MET higher baseline capacity an HR of 0.83 to 0.89, from an overview of 199 cohorts and 20.9 million observations with certainty graded very low to moderate (S1), consistent with the 2009 pooled RR of 0.87 per MET in healthy adults (S2) and with a referral cohort of 122,007 treadmill-tested patients in which the gradient had no upper limit (S3). That coefficient belongs to a trait, fitness, which integrates genotype, body composition, prior disease and training history. The intervention coefficient is a different and smaller quantity: randomized training in 464 sedentary postmenopausal women raised VO2peak by 4.2, 6.0 and 8.2 percent over six months at rising energy doses (S5), and a two-sample Mendelian randomization study with a VO2max instrument from about 70,000 adults found genetically predicted VO2max had no causal association with longevity or type 2 diabetes once body fat, lean mass and activity were in the model, while those upstream inputs did move VO2max (S4). Law: a per-MET mortality coefficient from a fitness cohort is a prognostic quantity and may not be multiplied by a training-induced MET gain to estimate a mortality effect; the two estimands share a unit and nothing else. Boundary: general-adult and referral cohorts for the gradient; sedentary women for the training gain; the diabetes limb of the MR is contested by a second MR, the longevity limb is not contradicted by any opened source.

The physics underneath

VO2max is the Fick product of cardiac output and arteriovenous oxygen difference, so a fitness test reads whole-organism oxygen delivery and extraction, which occult disease, low lean mass and high fat mass depress as surely as training raises it. The trait therefore carries illness and composition signal that a training assignment does not manipulate. Genetic instruments, randomized at conception, separate the two and find the upstream inputs causal and the scoreboard not.

The sources
← Previous axiom: 2.8, Consumer sleep staging detects sleep and misses wake, classifies stages at fair-to-moderate agreement, and averaging nights sharpens a biased stage estimate rather than correcting it; duration and timing are upheld, nightly deep and REM minutes are refutedNext axiom: 3.2, The activity-mortality curve is concave in log-dose →Link to this axiom: #ax-3-1
3.2 The activity-mortality curve is concave in log-doseVERIFIEDThe first dose buys most of the association and no harmful upper threshold survives reverse-causation correction.Open this axiomClose this axiom
The axiom in full

In pooled self-report cohorts (6 cohorts, 661,137 adults, median 14.2 years, 116,686 deaths) all-cause mortality HRs versus no leisure activity were 0.80 below 7.5 MET-h/week, 0.69 at 1 to 2 times the guideline minimum, 0.63 at 2 to 3 times, 0.61 at 3 to 5 times and 0.69 at 10 or more times, with no evidence of harm at the top (S6); a harmonised meta-analysis of non-occupational activity found RR 0.77 at about 75 min/week, 0.69 at 150 and 0.66 at 300 (S7); device-measured quartiles gave HRs 1.00, 0.48, 0.34, 0.27 with the same non-linear shape (S9). Fitting the three Arem bins against log-dose gives roughly a 10 percent lower hazard per doubling of weekly MET-hours (the anthropic leg's arithmetic on S6's published HRs, not the authors' claim). The apparent J or U shape at high volumes is a measurement artifact: in 133,819 health professionals with biennial activity reports, single-baseline or most-recent measures produced a J curve, while cumulative-average activity with a 2-year lag produced a monotone inverse association across 0 to 150 MET-h/week, and the per 10 MET-h/week HR moved from 0.78 (most recent) to 0.87 (cumulative average), so about half the apparent strength of the recent-activity estimate is reverse causation (S8). Law: the mortality association of activity is steepest at the inactive end, flattens in log-dose rather than in minutes, and once prodromal illness is lagged out shows diminishing returns without a demonstrated harmful ceiling in the achievable range. Boundary: middle-aged and older adult cohorts, self-report and device exposure, all-cause mortality, correlational class; the pooled associations are graded only suggestive in an umbrella review (S35).

The physics underneath

Each doubling of training load adds a roughly constant increment of physiological reserve (capillary density, stroke volume, insulin sensitivity) while the absolute minutes required double, so the hazard falls with the logarithm of dose. Illness suppresses recent activity before it causes death, which stacks the lowest and recently-reduced bins with dying people and bends the curve; lagging exposure ahead of the risk window removes the bend.

The sources
← Previous axiom: 3.1, Fitness predicts death; training moves fitness; the two coefficients are not the same numberNext axiom: 3.3, The size of the activity effect is a property of the exposure instrument: device measurement roughly doubles the apparent association and no unbiased estimate exists →Link to this axiom: #ax-3-2
3.3 The size of the activity effect is a property of the exposure instrument: device measurement roughly doubles the apparent association and no unbiased estimate existsVERIFIEDSelf-report error flattens the activity-mortality slope and device windows import illness signal that steepens it, so the measured effect size is set by the instrument.Open this axiomClose this axiom
The axiom in full

Self-reported leisure activity in pooled cohorts gives a maximum all-cause mortality reduction of about 30 to 40 percent (HR 0.61 to 0.69 at the plateau in 661,137 adults, S6; RR 0.66 at 300 min/week, S7), while accelerometer-measured total activity in a harmonised meta-analysis of 8 cohorts (n=36,383, mean age 62.6, median 5.8 years, 2,149 deaths) gives HR 0.27 (0.23 to 0.32) for the most active versus least active quarter, with light activity alone at HR 0.38 and MVPA at 0.52 (S9). The device estimate is not the truth either: a 7-day wear window captures current activity, which prodromal illness depresses, and the lag analysis in 133,819 adults shows recent-activity measures carry the largest reverse-causation inflation (per 10 MET-h/week HR 0.78 most-recent versus 0.87 cumulative-average, S8). Self-report dilutes the slope through misclassification; device measurement sharpens it and imports illness signal; early-event exclusion attenuates both (S7). Law: the measured size of the activity-mortality association varies about twofold with the exposure instrument and the lag applied, so a single hazard ratio for exercise is an instrument-and-design statement, never a biological constant, and the true causal effect plausibly lies between half the observational estimate and the estimate itself. Boundary: middle-aged and older adult cohorts, all-cause mortality, correlational class; no opened source measures dilution and reverse causation in the same cohort.

The physics underneath

Regression dilution: random error in a self-reported exposure flattens the fitted slope toward null. Reverse causation: a 7-day device window measures the organism's current state, and a dying organism moves less, which steepens the slope beyond the causal effect. The two biases pull in opposite directions and neither has been removed in the same dataset.

The sources
← Previous axiom: 3.2, The activity-mortality curve is concave in log-doseNext axiom: 3.4, Randomized exercise has demonstrated a healthspan effect (mobility disability) and has not demonstrated a mortality effect, and the gap is structural, set by event counts →Link to this axiom: #ax-3-3
3.4 Randomized exercise has demonstrated a healthspan effect (mobility disability) and has not demonstrated a mortality effect, and the gap is structural, set by event countsVERIFIEDDeaths are rare and disability common in healthy older volunteers, so randomized exercise resolves function and not mortality regardless of its true lifespan effect.Open this axiomClose this axiom
The axiom in full

In the LIFE trial, 1,635 sedentary adults aged 70 to 89 with functional limitation randomized to structured moderate activity versus health education had major mobility disability (inability to walk 400 m) in 30.1 versus 35.5 percent over a mean 2.6 years, HR 0.82 (0.69 to 0.98), and persistent disability HR 0.72 (0.57 to 0.91) (S12). In Generation 100, 1,567 Norwegian adults aged 70 to 77 randomized to five years of HIIT, moderate continuous training or guideline advice showed no all-cause mortality difference between control (4.7 percent) and the combined exercise arms, with HIIT HR 0.63 (0.33 to 1.20) and MICT HR 1.24 (0.73 to 2.10), controls who exercised at an intensity between the arms, 87.5 percent good baseline health and five-year survival above the national average (S10, S11). Lifestyle arms of diabetes-prevention trials did not move all-cause mortality at 21 years (DPPOS lifestyle HR 1.02, 0.81 to 1.28, 453 deaths, S14) or the cardiovascular composite at 9.6 years (Look AHEAD HR 0.95, 0.83 to 1.09, S15); the one long-run lifestyle mortality result, Da Qing's pooled diet and exercise arms at 30 years, HR 0.74 (0.61 to 0.89), does not isolate exercise (S16). Across the randomized literature 14,716 participants have been randomized to exercise in 57 mortality trials against more than 300,000 in drug trials (S19). Law: a trial of healthy volunteers accrues too few deaths to detect the 10 to 20 percent hazard difference cohorts predict for a realistic dose contrast (about 70 deaths in 1,567 people over five years detects a halving, not a fifth), while disability endpoints accrue hundreds of events in the same population, so exercise's demonstrated causal domain is function and its mortality effect is estimated from cohorts, not shown in trials. Boundary: older adults in trials; function and mortality endpoints; intervention class.

The physics underneath

Statistical power scales with the number of events, not participants. Healthy volunteers who can be randomized to exercise have low baseline hazard, and exercise cannot be blinded or placebo-controlled, so the control arm exercises too and the contrast shrinks. Mobility loss is common and fast in the frail-but-ambulatory, death is rare and slow, so the same trial resolves one endpoint and not the other.

The sources
← Previous axiom: 3.3, The size of the activity effect is a property of the exposure instrument: device measurement roughly doubles the apparent association and no unbiased estimate existsNext axiom: 3.5, Zone 2 is a non-portable label whose unique mortality benefit is unproven in both directions; the one randomized intensity contrast with a mortality endpoint runs against the moderate arm →Link to this axiom: #ax-3-4
3.5 Zone 2 is a non-portable label whose unique mortality benefit is unproven in both directions; the one randomized intensity contrast with a mortality endpoint runs against the moderate armVERIFIEDAdaptation signaling scales with intensity and mortality cohorts measure only volume, so no evidence isolates a unique benefit for sub-threshold continuous work.Open this axiomClose this axiom
The axiom in full

The label does not denote one exposure: Seiler's three-zone model places zone 2 between the first and second ventilatory thresholds in elite endurance athletes (S33), while popular use places it below the first threshold, so a heart-rate percentage, a lactate band and a threshold-anchored domain are different interventions. On adaptation, a within-subject single-leg trial in 10 young men found six sessions of interval training over two weeks produced greater increases in citrate synthase activity and mitochondrial respiration than continuous training matched for total work and duration (S34), and a 2025 narrative review reports AMPK and PGC-1alpha activation scaling with intensity and small, inconsistent activation after moderate continuous work (S32), so the claim that sub-threshold work uniquely maximizes mitochondrial adaptation is contradicted at the signaling level in young men. On mortality, self-report cohorts find the guideline minimum met by moderate or vigorous activity associated with near-maximal benefit at matched MET-hours (S6), device cohorts find any intensity associated with lower mortality (S9), and the only randomized intensity contrast with all-cause mortality as the declared primary endpoint, Generation 100, had the moderate continuous arm (about 70 percent of peak heart rate) numerically worst of three (5.9 percent deaths; MICT HR 1.24, 0.73 to 2.10 versus control; HIIT versus MICT HR 0.51, 0.25 to 1.02), underpowered and with a contaminated control (S10, S11). Law: no opened source isolates a sub-threshold continuous prescription and shows a mortality advantage; total dose, not its intensity distribution, carries the verified association; and the single randomized datum is unfavorable to the moderate arm without refuting it. Boundary: athletes for the definition, young men for the adaptation result, older adults for the trial, adult cohorts for the associations; blog-sourced fold-change figures quoted by one leg are excluded.

The physics underneath

Mitochondrial biogenesis is triggered by cellular energy stress (ATP turnover, AMP accumulation, calcium flux) whose magnitude rises with intensity; low-intensity work earns its place through low fatigue cost per minute, which lets volume accumulate, not through a privileged signal. Mortality cohorts measure volume in MET-hours and cannot resolve an intensity-distribution hypothesis.

The sources
← Previous axiom: 3.4, Randomized exercise has demonstrated a healthspan effect (mobility disability) and has not demonstrated a mortality effect, and the gap is structural, set by event countsNext axiom: 3.6, Strength is two different variables →Link to this axiom: #ax-3-5
3.6 Strength is two different variablesVERIFIEDA grip-strength gradient that reads illness and nutrition as much as muscle, and a trainable capacity whose mortality effect has never been randomized.Open this axiomClose this axiom
The axiom in full

In 42 cohort studies of 3,002,203 community-dwelling adults, lowest versus highest grip strength carried all-cause mortality HR 1.41 (1.30 to 1.52) and each 5 kg decrease HR 1.16 (1.12 to 1.20), linear to 56 kg and surviving exclusion of baseline cardiovascular disease and cancer (S27); within PURE's 155,722 adults, low grip strength carried a mortality population attributable fraction of 11.6 percent (7.3 to 16.0) against 2.2 percent for low physical activity (S22). Muscle-strengthening activity, the behavior, is associated with 10 to 17 percent lower all-cause mortality in 16 cohorts, J-shaped with the maximum around 30 to 60 minutes a week and unclear at higher volumes (S26). On the intervention side, progressive resistance training in 121 trials of older adults produced a large effect on strength and a smaller effect on physical function (S28), protein supplementation augments those strength and lean-mass gains with no mortality endpoint (S29), combined training adds CRF and lean mass over aerobic training alone in CAD patients without changing biomarkers (S39), and the anthropic leg's extract reports grip strength predicting mortality more strongly (HR 0.69) than the more trainable, more load-bearing knee extension (HR 0.86) in about 2 million adults, the tell that the grip gradient carries disease signal. No resistance-training trial with all-cause mortality as a primary endpoint was found by any leg. Law: the grip gradient is a marker of neuromuscular integrity, nutrition and occult disease and cannot be converted into a lifting prescription; the behavior's association is modest and J-shaped in minutes; the only demonstrated causal effects of resistance training are on strength, lean mass and, less strongly, function. Boundary: community-dwelling and older adults; mortality is correlational, strength and function are intervention class.

The physics underneath

Grip force is cheap to measure and falls with sarcopenia, catabolic illness, inflammation and malnutrition, which is why it predicts death well and why the prediction is not a training target. Resistance training raises force through neural recruitment and hypertrophy; transfer to function is partial because function depends on power, balance and task demand (power equals force times velocity), so strength gains outrun functional gains.

The sources
← Previous axiom: 3.5, Zone 2 is a non-portable label whose unique mortality benefit is unproven in both directions; the one randomized intensity contrast with a mortality endpoint runs against the moderate armNext axiom: 3.7, Levers may be ranked only inside one design, population, endpoint and follow-up; where those do not match, noncomparability is the result, and the few matched comparisons show exercise and drugs trading places by condition →Link to this axiom: #ax-3-6
3.7 Levers may be ranked only inside one design, population, endpoint and follow-up; where those do not match, noncomparability is the result, and the few matched comparisons show exercise and drugs trading places by conditionVERIFIEDA hazard ratio carries its own population's baseline risk, design and follow-up, so ratios from unmatched studies cannot be ranked, only described side by side.Open this axiomClose this axiom
The axiom in full

A cross-lever comparison requires the same endpoint, the same population and baseline hazard, the same follow-up, a commensurable exposure contrast and the same evidence class. The matched exceptions that exist: in a network meta-analysis of 305 RCTs with 339,274 participants (14,716 in 57 exercise trials), exercise and drug interventions showed no detectable mortality difference in secondary prevention of coronary heart disease and in prediabetes, exercise beat anticoagulants (OR 0.09) and antiplatelets (OR 0.10) after stroke, and diuretics beat exercise in heart failure (OR 4.11) (S19); in the DPP, lifestyle cut incident diabetes 58 percent and metformin 31 percent in the same 3,234 adults over 2.8 years, and neither moved all-cause mortality over 21 years (S13, S14). The unmatched pairs that are routinely ranked anyway: an exercise cohort RR of 0.69 at 150 min/week (S7) against SPRINT's randomized all-cause mortality HR 0.73 in 9,361 hypertensive adults over 3.26 years (S18), SELECT's HR 0.81 in 17,604 adults with established cardiovascular disease over 34 months (S20), or statins' RR 0.90 against placebo across 55 trials (S21); and the SPAN cohort's HR 0.36 for optimal sleep, activity and diet tertiles in 59,078 adults over 8.1 years (S30) against any drug trial. Those pairs differ in design, baseline hazard (several-fold), follow-up, blinding and contrast, so the drug's randomized demonstration is a statement about its trial and not about biology ranking above exercise, and the cohort's larger ratio is a statement about healthy-volunteer selection and compounding habits and not about biology ranking above the drug; within SPRINT itself survival improved while gait speed and mobility did not (S37), and in PREDIMED a cardiovascular composite HR of about 0.70 sat beside an all-cause mortality HR of 0.98 (0.77 to 1.24) in the same older high-risk Spanish adults (S17), so even endpoints inside one trial are not interchangeable. Law: a demonstrated effect and a large association are different quantities; two levers are comparable only when all five conditions hold, and absent that the honest output is a joint description with the mismatches named. Boundary: secondary-prevention and high-risk populations for the matched RCT comparisons; healthy cohorts for the associations.

The physics underneath

A hazard ratio is a within-study contrast between two arms of one population under one design; it carries that population's baseline hazard, that design's biases and that follow-up's time-to-effect. Transporting it beside another study's ratio multiplies the two sets of unmatched terms. Pharmacology buys randomized mortality demonstrations with high-hazard populations, large samples and placebo blinding, none of which exercise can have.

The sources
← Previous axiom: 3.6, Strength is two different variablesNext axiom: 3.8, On within-framework burden ledgers the non-exotic exposures carry a large but mediation-sensitive share of mortality, with diet, blood pressure, smoking and glucose above inactivity, and the levers compound in cohorts without a demonstrated additive law →Link to this axiom: #ax-3-7
3.8 On within-framework burden ledgers the non-exotic exposures carry a large but mediation-sensitive share of mortality, with diet, blood pressure, smoking and glucose above inactivity, and the levers compound in cohorts without a demonstrated additive lawVERIFIEDAttributable share equals risk times prevalence under a chosen counterfactual, so inactivity's share spans 1 to 16 percent while diet and blood pressure outrank it on every ledger.Open this axiomClose this axiom
The axiom in full

The only defensible cross-lever share comparisons are within one framework. In PURE (155,722 adults, 21 countries, median about 9.5 years) the mortality population attributable fractions were poor diet 11.1 percent (7.7 to 14.6), low grip strength 11.6 percent (7.3 to 16.0), hypertension 8.8 percent (7.6 to 9.9) and low physical activity 2.2 percent (1.0 to 3.3), with sleep not assessed (S22). In GBD 2021, dietary risks were attributed about 8 million deaths, high fasting plasma glucose 5.4 million and low physical activity 0.686 million (ranked 14th), with high systolic blood pressure, smoking and glucose the top three risks (S23); GBD 2019 attributed 2.4 million deaths to alcohol and 10.8 million to high systolic blood pressure and treated sugar-sweetened beverages, not sugar, as the dietary component (S24). A different counterfactual gives a different number for the same exposure: Garcia's potential impact fraction of 15.7 percent of deaths if all cohort members met the activity minimum, assuming causality (S7), against GBD's roughly 1 percent, with a rival 7.2 percent all-cause share for inactivity reported in the anthropic leg, a spread driven by how mediation through blood pressure, adiposity and glucose is assigned and by GBD version instability (S25). Combined-lever cohorts show compounding: in UK Biobank SPAN, optimal joint tertiles of device-measured sleep (7.2 to 8.0 h), MVPA (42 to 103 min/day) and diet quality carried HR 0.36 (0.26 to 0.50) versus the lowest tertiles, and a minimum combined shift of 15 min sleep, 1.6 min MVPA and 5 diet points HR 0.90 (0.88 to 0.93) (S30), with about 9.35 lifespan-years and 9.45 healthspan-years between jointly favourable and unfavourable tertiles (S31); joint activity-and-sleep associations are nonlinear and interacting (S38), and sleep alone enters the ledgers only as a correlational U-curve versus 7 hours (RR 1.07, 1.04, 1.01 at 4, 5, 6 h; 1.07, 1.21, 1.37, 1.55 at 8 to 11 h) whose claims The Physics of Sleep holds and the incremental sleep healthspan evidence vector sizes (S36). Law: the share of mortality the non-exotic exposures carry is a modeled quantity that ranges several-fold with the counterfactual and the mediation rule; within any one ledger inactivity sits below diet, blood pressure, smoking and glycemia; and the joint cohort association of the levers is far larger than any single lever's, with no demonstrated additive or causal decomposition. Boundary: adult cohorts and global burden models; attributable fractions and hazard ratios are correlational and model class, not trial effects; the ledgers test sugar-sweetened beverages and heavy drinking, not 'no sugar' or 'no alcohol'.

The physics underneath

A population attributable fraction multiplies relative risk by exposure prevalence under a chosen counterfactual, so a common exposure with a modest risk ratio (poor diet) outranks a rarer or weaker one (inactivity), and a hierarchical model that routes activity's effect through blood pressure and glucose credits those downstream risks. People who sleep, move and eat well cluster on wealth, health and absence of illness, so joint-tertile contrasts in cohorts capture compounding habits and healthy-volunteer selection together.

The sources
← Previous axiom: 3.7, Levers may be ranked only inside one design, population, endpoint and follow-up; where those do not match, noncomparability is the result, and the few matched comparisons show exercise and drugs trading places by conditionNext axiom: 4.1, A1. Sustained energy surplus, stored past subcutaneous capacity, is the master nutrition lever; it moves the whole cardiometabolic panel in randomized trials and the other three exposures act mostly through it →Link to this axiom: #ax-3-8

How do protein, added sugar, alcohol and energy balance act on healthspan through adiposity, insulin resistance and lipids, and does glucose variability predict or improve outcomes in non-diabetics?

Nutrition metabolic health and glucose · eight axioms

The mechanism under test is nutrition and metabolic health: protein, sugar, alcohol and energy balance acting through adiposity, insulin resistance and lipids, and what glucose variability predicts in non-diabetics. The easy answer this vector was told not to settle for: “No sugar and a flat glucose trace prove metabolic health”
4.1 A1. Sustained energy surplus, stored past subcutaneous capacity, is the master nutrition lever; it moves the whole cardiometabolic panel in randomized trials and the other three exposures act mostly through itVERIFIEDEnergy stored past subcutaneous capacity spills into liver and muscle, blunting insulin action and raising VLDL output; cutting the surplus reverses that panel first.Open this axiomClose this axiom
The axiom in full

In healthy non-obese adults aged 21 to 50 (CALERIE 2, 218 randomized, 2 years, INTERVENTION), a prescribed 25 percent calorie restriction that achieved 11.9 percent produced about 10 percent weight loss (71 percent of it fat) and lowered LDL cholesterol, total-to-HDL ratio, systolic and diastolic blood pressure, C-reactive protein and improved an insulin-sensitivity index; the panel changes held after adjustment for the weight change itself. In adults with obesity (Magkos 2016, 40 randomized to weight maintenance or progressive loss, INTERVENTION), about 5 percent weight loss improved insulin sensitivity in adipose tissue, liver and muscle together and cut intrahepatic fat, with further gains at 11 and 16 percent. The surplus itself is set by food form more than by label arithmetic: in 20 inpatient adults fed ultra-processed and unprocessed diets matched for presented calories, sugar, fat, sodium, fibre and macronutrients for 14 days each in random order (Hall 2019, INTERVENTION crossover), ad libitum intake ran about 500 kcal/day higher on the ultra-processed diet (+280 from carbohydrate, +230 from fat, -2 from protein) and weight moved +0.9 versus -0.9 kg, correlation of weight change with energy intake r = 0.8, meals eaten at 48 versus 31 kcal/min. Boundary: the CALERIE population was non-obese and young-to-middle-aged; the Magkos population had obesity; the Hall crossover was 14 days per arm in a ward. The mechanism from surplus to insulin resistance (subcutaneous overflow, ectopic lipid, diacylglycerol-activated kinase blockade of insulin signalling) is MODEL and review grade in these legs, not measured here.

The physics underneath

Fat mass is the time integral of intake minus expenditure (conservation, an accounting identity that explains nothing about why either term changed). Once subcutaneous adipose storage is exceeded, lipid is deposited in liver, muscle and pancreas; ectopic lipid intermediates impair insulin signalling and raise hepatic VLDL output; this is the shared trunk through which protein surplus, liquid sugar and ethanol reach lipids and, last, glycaemia. Reversing the surplus reverses the ectopic store first, which is why 5 percent loss already moves three tissues' insulin sensitivity.

The sources
← Previous axiom: 3.8, On within-framework burden ledgers the non-exotic exposures carry a large but mediation-sensitive share of mortality, with diet, blood pressure, smoking and glucose above inactivity, and the levers compound in cohorts without a demonstrated additive lawNext axiom: 4.2, A2. Every nutrient effect is a substitution effect at a stated energy level; the same gram of protein or sugar added, swapped or removed yields different outcomes, so a lever named without its replacement is not yet a claim →Link to this axiom: #ax-4-1
4.2 A2. Every nutrient effect is a substitution effect at a stated energy level; the same gram of protein or sugar added, swapped or removed yields different outcomes, so a lever named without its replacement is not yet a claimVERIFIEDA nutrient enters a diet only by displacing another or adding energy, so its measured effect is the pair's effect; the displaced term carries part of the cause.Open this axiomClose this axiom
The axiom in full

In OmniHeart (164 adults with prehypertension or stage 1 hypertension, three controlled-feeding diets in randomized crossover at stable weight, INTERVENTION, 2005), replacing about 10 percent of energy from carbohydrate with protein lowered LDL cholesterol and triglycerides and blood pressure, and replacing it with unsaturated fat lowered blood pressure and triglycerides and raised HDL, all without weight change. In Bray 2012 (25 healthy adults overfed about 40 percent above requirement for 8 weeks at 5, 15 or 25 percent protein, INTERVENTION), fat gain was the same across arms (about 3.5 kg) while lean mass and energy expenditure rose only in the 15 and 25 percent arms and fell in the 5 percent arm; the scale would have misread the result. In Te Morenga 2012 (systematic review of trials in adults, INTERVENTION synthesis), reducing dietary sugars in ad libitum diets lowered body weight by 0.80 kg (95 percent CI 0.39 to 1.21, 5 trials) while isoenergetic exchange of sugars for other carbohydrates produced no weight change. Boundary: these are weeks-to-months surrogate endpoints (lipids, blood pressure, body composition), in prehypertensive, healthy or overweight adults; none is a hard-outcome trial.

The physics underneath

At fixed energy, raising one macronutrient lowers another, so the exposure is always a pair (what came in, what left); the lipid and pressure responses in OmniHeart arise from the displaced carbohydrate as much as from the added protein or fat. Under surplus, protein's amino acids are partitioned to lean tissue and to thermogenesis rather than sparing fat storage, which is why fat gain matched across arms. Under ad libitum eating, removing sugar works mainly by removing energy; under isoenergetic exchange it does not, which separates the energy channel from any composition channel.

The sources
← Previous axiom: 4.1, A1. Sustained energy surplus, stored past subcutaneous capacity, is the master nutrition lever; it moves the whole cardiometabolic panel in randomized trials and the other three exposures act mostly through itNext axiom: 4.3, A3. Protein is a threshold nutrient with a plateau, not a dose lever: above about 1.6 g/kg/day in resistance-trained adults extra protein adds no further lean mass, the whole supplementation effect is fractions of a kilogram, and no trial links protein above adequacy to longer life →Link to this axiom: #ax-4-2
4.3 A3. Protein is a threshold nutrient with a plateau, not a dose lever: above about 1.6 g/kg/day in resistance-trained adults extra protein adds no further lean mass, the whole supplementation effect is fractions of a kilogram, and no trial links protein above adequacy to longer lifeVERIFIEDLoading, not substrate, sets muscle synthesis; past the saturating protein dose amino acids are oxidized, so lean-mass gain plateaus near 1.6 g/kg/day in trained adults.Open this axiomClose this axiom
The axiom in full

In Morton 2018 (meta-analysis and meta-regression of 49 randomized trials, 1,863 healthy adults doing resistance training, INTERVENTION synthesis), protein supplementation added about 0.30 kg of fat-free mass (and 2.49 kg to one-repetition maximum) over training alone, and the regression breakpoint beyond which more protein added no fat-free mass sat at 1.62 g/kg/day with a 95 percent confidence interval reaching about 2.2 g/kg/day; the effect fell with age and rose with training experience. In an acute feeding meta-analysis (Wolever 2024, controlled feeding trials in adults without diabetes, INTERVENTION synthesis), adding protein to a carbohydrate meal lowered the postprandial glucose area and raised the insulin area, a glucose-lowering mechanism that is not evidence of improved insulin sensitivity. The only outcome-adjacent evidence against surplus is CORRELATIONAL: in NHANES III (Levine 2014, 6,381 adults 50 and over followed about 18 years), adults aged 50 to 65 in the high-protein group (20 percent or more of energy) had a hazard ratio of 4.33 for cancer mortality versus the low-protein group (under 10 percent), and the association reversed after 65, where higher protein tracked lower mortality; this is one cohort, one diet recall, and it has not been reproduced by a trial. Boundary: the plateau is measured in trained adults in energy balance; the requirement rises in energy deficit and in older adults with anabolic resistance; no trial measures protein against all-cause mortality or disability.

The physics underneath

Muscle protein synthesis saturates with amino-acid delivery because the stimulus is mechanical loading, not substrate; above the saturating dose, amino acids are deaminated and oxidized or converted to glucose, so surplus is an occupied energy slot rather than more tissue. Requirement-type nutrients therefore show a steep response below adequacy and a flat one above, the opposite shape from a toxin.

The sources
← Previous axiom: 4.2, A2. Every nutrient effect is a substitution effect at a stated energy level; the same gram of protein or sugar added, swapped or removed yields different outcomes, so a lever named without its replacement is not yet a claimNext axiom: 4.4, A4. The exposure behind "no sugar" is hepatic fructose load, chiefly from sweetened drinks: at matched energy fructose and sucrose raise liver fat synthesis and glucose does not, and removing the drink yields liver-fat loss concentrated in people who drink a lot and already carry liver fat, with little weight change →Link to this axiom: #ax-4-3
4.4 A4. The exposure behind "no sugar" is hepatic fructose load, chiefly from sweetened drinks: at matched energy fructose and sucrose raise liver fat synthesis and glucose does not, and removing the drink yields liver-fat loss concentrated in people who drink a lot and already carry liver fat, with little weight changeVERIFIEDFructose bypasses the liver's glycolytic feedback control, so sweetened-drink fructose at matched energy drives de novo lipogenesis and liver fat; glucose at the same dose does not.Open this axiomClose this axiom
The axiom in full

In Geidl-Flueck 2021 (94 healthy men, randomized to 80 g/day of fructose-, sucrose- or glucose-sweetened beverage or abstinence for 7 weeks, INTERVENTION), basal hepatic fatty-acid synthesis roughly doubled in the fructose and sucrose arms and did not change in the glucose arm, with reported total energy intake similar across arms; this is a flux measurement, not a liver-fat stock or an outcome. In Maersk 2012 (47 overweight adults randomized to 1 L/day of sucrose cola, isocaloric semi-skim milk, aspartame cola or water for 6 months, INTERVENTION, magnetic resonance spectroscopy), the sucrose drink raised liver, muscle and visceral fat relative to the isocaloric milk, so the form of the energy, not its amount, set the depot. In REDUCS (Campos 2015, 31 overweight adults drinking at least 660 mL/day of sugar-sweetened beverage, 12 weeks of replacement with artificially sweetened drinks, INTERVENTION), intrahepatocellular lipid fell to 74 percent of baseline, the fall concentrated in those above 60 mmol/L at baseline, with no change in body weight, visceral fat or other metabolic markers. "Added sugars" is a labelling category (sugars added in processing plus syrups, honey and concentrated juices, per the FDA label rule), not a molecule the liver detects; the biochemical category is fructose fraction and delivery rate. Boundary: healthy or overweight adults, liquid sugar at 80 g/day or about 1 L/day, 7 weeks to 6 months, liver-fat and lipid-flux endpoints; the opened legs carried no trial of added sugar against a hard outcome, and the observational link runs through ultra-processed food (an umbrella review of 39 meta-analyses rates diabetes and obesity associations as highly suggestive, CORRELATIONAL), which does not isolate sugar from energy density and matrix.

The physics underneath

Fructose is cleared on first pass by the liver and enters glycolysis below the phosphofructokinase control point, so its flux is not fed back by cellular energy state; the carbon is pushed to de novo lipogenesis and VLDL export or stored as liver fat. Glucose at the same dose is regulated and distributed to muscle. Liquid delivery raises the rate, and the lipogenic response steepens where hepatic oxidative capacity is exceeded, which is why benefit concentrates in high-dose consumers with existing steatosis and is near zero at low baseline intake.

The sources
← Previous axiom: 4.3, A3. Protein is a threshold nutrient with a plateau, not a dose lever: above about 1.6 g/kg/day in resistance-trained adults extra protein adds no further lean mass, the whole supplementation effect is fractions of a kilogram, and no trial links protein above adequacy to longer lifeNext axiom: 4.5, A5. Alcohol's J-shaped mortality curve is a reference-group artifact: with former and occasional drinkers out of the abstainer group, low-volume drinking shows no mortality protection, risk rises monotonically from about 25 g/day, and genetic instruments show harm to blood pressure and stroke with no cardioprotection; the absolute gradient at low dose is shallow and no hard-endpoint trial exists →Link to this axiom: #ax-4-4
4.5 A5. Alcohol's J-shaped mortality curve is a reference-group artifact: with former and occasional drinkers out of the abstainer group, low-volume drinking shows no mortality protection, risk rises monotonically from about 25 g/day, and genetic instruments show harm to blood pressure and stroke with no cardioprotection; the absolute gradient at low dose is shallow and no hard-endpoint trial existsVERIFIEDSick quitters in the abstainer group and survivors among drinkers manufacture the J; cleaned cohorts and genetic instruments show no protection and dose-rising harm.Open this axiomClose this axiom
The axiom in full

In Zhao 2023 (107 cohort studies, 4,838,825 participants, 425,564 deaths, CORRELATIONAL synthesis with prespecified bias adjustments), the relative risk of all-cause mortality for 1.3 to under 25 g/day of ethanol versus lifetime non-drinkers was 0.93 (95 percent CI 0.85 to 1.01), and for 45 to under 65 g/day 1.19 (1.07 to 1.32), rising further above that. In Stockwell 2024 (the same 107 cohorts analysed by study quality), cohorts with mean age 55 or under, follow-up past 55 and former and occasional drinkers removed from the reference group gave RR 0.98 (0.87 to 1.11) for low-volume drinkers, while lower-quality cohorts gave 0.84 (0.79 to 0.89); among non-smoking cohorts the mean low-volume RR was 1.16 (0.91 to 1.41). The Sarich 2024 umbrella review of reviews using lifetime abstainers or low-volume drinkers as reference reached the same shape. A MESA re-analysis (CORRELATIONAL) found that removing former drinkers fixes the abstainer side but not survivorship among current drinkers, and correcting both shifted light drinking toward harm and gave a dose response with coronary artery calcium. In the China Kadoorie Biobank (Millwood 2019, about 512,000 adults, Mendelian randomization on ALDH2 and ADH1B variants, MODEL-grade instrument), genotype-predicted alcohol intake raised blood pressure and stroke risk with no protective range, and gave no clear effect on myocardial infarction. Boundary: mortality and cardiovascular endpoints in middle-aged and older cohorts; the genetic evidence is East Asian; no randomized trial of moderate drinking has been completed.

The physics underneath

Ethanol is a pharmacological exposure and an energy substrate at once: it is oxidized preferentially through acetaldehyde and acetate, shifting hepatic redox toward fat synthesis and away from fat oxidation, and it raises blood pressure; none of this has a dose below which the mechanism switches off. The apparent benefit arises in the data, not the body: sick quitters inflate the reference group's mortality, and survivors among heavy drinkers deflate the high-dose harm.

The sources
← Previous axiom: 4.4, A4. The exposure behind "no sugar" is hepatic fructose load, chiefly from sweetened drinks: at matched energy fructose and sucrose raise liver fat synthesis and glucose does not, and removing the drink yields liver-fat loss concentrated in people who drink a lot and already carry liver fat, with little weight changeNext axiom: 4.6, A6. Glycaemia is the last variable to move in the diet-to-disease chain, so a flat glucose trace does not certify metabolic health: within-person feeding trials move intake, lipids and body composition in directions the glucose trace does not show, and non-diabetic glucose variability tracks beta-cell function rather than insulin sensitivity, lipids or adiposity →Link to this axiom: #ax-4-5
4.6 A6. Glycaemia is the last variable to move in the diet-to-disease chain, so a flat glucose trace does not certify metabolic health: within-person feeding trials move intake, lipids and body composition in directions the glucose trace does not show, and non-diabetic glucose variability tracks beta-cell function rather than insulin sensitivity, lipids or adiposityVERIFIEDBeta-cell compensation holds glucose normal while liver fat, insulin resistance and apoB climb, so meal glucose is the last and most lagged readout of diet.Open this axiomClose this axiom
The axiom in full

In Hall 2021 (20 adults without diabetes, inpatient randomized crossover, two weeks each on a low-fat plant-based diet and a ketogenic diet, INTERVENTION), the higher-carbohydrate diet produced higher postprandial glucose and insulin yet participants ate about 689 kcal/day less on it than on the ketogenic diet, so the flatter trace sat on the diet with the higher spontaneous intake. In OmniHeart (164 adults, controlled feeding at stable weight, INTERVENTION) lipids and blood pressure moved under substitution with no weight change and no glucose endpoint driving it; in Bray 2012 (25 adults, INTERVENTION) fat gain was identical while the scale and lean mass diverged. In Hjort 2024 (systematic review of 71 studies of glucose variability in people without diabetes, mostly cross-sectional, CORRELATIONAL; searches to April 2022), variability was higher in prediabetes and associated with beta-cell dysfunction, while its relations to insulin sensitivity, adiposity, fatty liver, lipids and blood pressure were inconsistent or absent. Apolipoprotein B is prioritized as the lipid determinant of coronary disease in multivariable Mendelian randomization (MODEL-grade instrument), and the chain from surplus, liquid fructose and ethanol to atherosclerosis runs through liver fat, hepatic insulin resistance and VLDL-apoB output, not through meal glucose amplitude. The lipid rise under carbohydrate restriction in lean people ("lean mass hyper-responder") was carried by one leg from secondary pages only and rides here as a flag, not a measurement. Boundary: healthy, prehypertensive and obese adults; surrogate endpoints over days to months; the ordering of the chain is inferred from mechanism and from cross-sectional data, which is why this axiom claims PROBABLE and not FORGED.

The physics underneath

Beta-cell compensation holds glucose near normal for years while adiposity, ectopic liver fat, hepatic insulin resistance and particle output are already moving; glucose becomes abnormal only after compensation fails. An instrument reading the terminal node therefore has the longest lag and the weakest link to the levers; the proximal nodes (intake, liver fat, apoB, fasting insulin) are measurable earlier and move first.

The sources
← Previous axiom: 4.5, A5. Alcohol's J-shaped mortality curve is a reference-group artifact: with former and occasional drinkers out of the abstainer group, low-volume drinking shows no mortality protection, risk rises monotonically from about 25 g/day, and genetic instruments show harm to blood pressure and stroke with no cardioprotection; the absolute gradient at low dose is shallow and no hard-endpoint trial existsNext axiom: 4.7, A7. "Glucose variability predicts outcomes" is true for years-scale visit-to-visit fasting glucose variability and unproven for 14-day sensor variability in healthy adults; the two are different physical quantities sharing a word, and the one sensor-based prognostic result to date is a learned pattern in a mixed cohort, not an isolated variability effect →Link to this axiom: #ax-4-6
4.7 A7. "Glucose variability predicts outcomes" is true for years-scale visit-to-visit fasting glucose variability and unproven for 14-day sensor variability in healthy adults; the two are different physical quantities sharing a word, and the one sensor-based prognostic result to date is a learned pattern in a mixed cohort, not an isolated variability effectVERIFIEDYears-scale fasting-glucose variability indexes a drifting person and forecasts outcomes; a fortnight's sensor variability indexes meals, and no cohort ties it to healthy-adult outcomes.Open this axiomClose this axiom
The axiom in full

The outcome evidence attaches to slow variability measured in laboratories across years: in CARDIA (young adults followed about 30 years, CORRELATIONAL), each 1 SD higher average real variability of fasting glucose before any diabetes carried hazard ratios of 1.64 for incident diabetes, 1.15 for cardiovascular disease and 1.25 for all-cause mortality after adjustment; in a 2017 cohort of adults without diabetes (CORRELATIONAL), the highest versus lowest quartile of visit-to-visit fasting glucose variability carried a hazard ratio of 1.55 for all-cause mortality; in a Korean national cohort of about 3.2 million adults (CORRELATIONAL), the highest variability quartile carried hazard ratios of 1.08 for myocardial infarction, 1.09 for stroke and 1.12 for death. A 14-day sensor coefficient of variation is a fast response variable set by meal composition, timing and the preceding exercise bout; in Hjort 2024 (71 studies, CORRELATIONAL) non-diabetic sensor variability tracked beta-cell function and was not consistently related to the other risk factors, and the leg that searched found no cohort linking sensor-derived variability in healthy adults to hard outcomes. The nearest result is GluFormer (published 14 January 2026, MODEL): a representation learned from sensor traces of 10,812 adults, mostly without diabetes, predicted incident diabetes and cardiovascular mortality in the AEGIS cohort (580 adults, median 11 years of follow-up) better than HbA1c, with most events in the top predicted-risk quartile; it evaluates a learned pattern in a cohort not shown to be uniformly normoglycaemic, not the causal weight of variability itself. In ZOE PREDICT (adults without diabetes, CORRELATIONAL, company-funded preprint) sensor variability associated with diet and lifestyle; in 53 non-diabetic and 52 pre-diabetic Indian adults over 14 days (CORRELATIONAL) sensor metrics separated the two groups. Boundary: the hazard ratios belong to the laboratory estimand in middle-aged cohorts and transfer to no sensor metric; the sensor estimand has association and one prognostic model, no outcome cohort in healthy adults.

The physics underneath

Variability across annual fasting draws indexes a drifting state of the person (beta-cell reserve, weight trajectory, illness, medication) and so forecasts the person's trajectory; variability across a fortnight of meals indexes the meals and the day before. A forecasting variable can carry outcome signal while being causally inert, so prediction, where it exists, says nothing about whether suppressing the excursion helps.

The sources
← Previous axiom: 4.6, A6. Glycaemia is the last variable to move in the diet-to-disease chain, so a flat glucose trace does not certify metabolic health: within-person feeding trials move intake, lipids and body composition in directions the glucose trace does not show, and non-diabetic glucose variability tracks beta-cell function rather than insulin sensitivity, lipids or adiposityNext axiom: 4.8, A8. Monitor-guided eating has no outcome evidence in normoglycaemic adults: the randomized literature is a diabetes literature, the pooled effects are small and glycaemic, the normoglycaemic subgroup shows no glycaemic benefit, and where a sensor helps it is as an adherence channel inside a programme whose contact time was never matched →Link to this axiom: #ax-4-7
4.8 A8. Monitor-guided eating has no outcome evidence in normoglycaemic adults: the randomized literature is a diabetes literature, the pooled effects are small and glycaemic, the normoglycaemic subgroup shows no glycaemic benefit, and where a sensor helps it is as an adherence channel inside a programme whose contact time was never matchedVERIFIEDSensor feedback changes eating through salience, a channel any diary can fill; in normoglycaemic adults the target metric barely moves and no outcome trial exists.Open this axiomClose this axiom
The axiom in full

In Richardson 2024 (systematic review and meta-analysis of 25 randomized trials of sensor feedback for behaviour change, INTERVENTION synthesis), 17 trials were in type 2 diabetes, 3 in type 1, 3 in gestational diabetes and 3 in obesity; pooled feedback lowered HbA1c by 0.28 percentage points and raised time in range by 7.4 percentage points, with no significant effect on BMI and a pooled weight change of -0.7 kg across 8 studies (I-squared 0 percent); only 4 of 25 trials measured dietary change and 5 of 25 physical activity. A 2026 review restricted to people without diabetes (UNVERIFIED live fetch; the anthropic leg verified it at snippet level only) reported that sensor use may lower mean glucose versus control but did not change BMI or glycaemic variability, and that in subgroup analysis the glycaemic benefit appeared in prediabetes and not in normoglycaemic adults. In the ZOE METHOD trial (347 general adults, 18 weeks, INTERVENTION), a personalized programme that bundled sensor data with microbiome and other inputs, education and an app improved triglycerides, weight and HbA1c and did not change LDL; contact and intensity were not matched, so the sensor's share is unidentified. The one non-diabetic eating-pattern result opened is exploratory (17 postmenopausal women with obesity, 16 weeks, a secondary analysis). The Diabetes Prevention Program lowered incident diabetes by 58 percent over about three years in high-risk adults with a lifestyle programme and no sensor, which fixes the comparator any sensor claim must beat. The leg that searched found no trial comparing sensor feedback against food logging or daily weighing at matched contact. Boundary: adults with diabetes, prediabetes, obesity or general adults in programmes; glycaemic and weight surrogates over 12 to 24 weeks; no trial in normoglycaemic adults with a disease, disability or mortality endpoint, and no trial isolating glucose information from attention.

The physics underneath

A sensor changes behaviour through salience and immediacy of feedback, a channel any self-monitoring instrument can occupy; in a person whose glucose is already normal the feedback has little to correct, so the target metric barely moves and whatever benefit appears must travel through intake and adherence, not glycaemia. Identifying the sensor's own effect requires randomizing the information while holding the diet and the coaching fixed, which no opened trial did.

The sources
← Previous axiom: 4.7, A7. "Glucose variability predicts outcomes" is true for years-scale visit-to-visit fasting glucose variability and unproven for 14-day sensor variability in healthy adults; the two are different physical quantities sharing a word, and the one sensor-based prognostic result to date is a learned pattern in a mixed cohort, not an isolated variability effectThis is the last axiom on the page.Link to this axiom: #ax-4-8

The easy answers this run refused

Each vector was told the common saying it must not settle for, and told to rebuild the answer from first principles instead.

Aging runs on mechanisms that instruments can measure, and every mechanism on this page can be checked again.