Foundational
Large prospective cohorts (≥100k participants), meta-analyses that converge across populations, Mendelian-randomization causal evidence, or consistent RCTs. The things to prioritize — and where most of this page sits.
rec.003
doc: framework
evidence base: v1.2
domains: 5
composite score: none
The framework · evidence base v1.2
Five domains. Every recommendation carries its study and its strength of evidence.
This is what Ilfirin coaches toward. Not one expert's system — a synthesis of large prospective cohorts, meta-analyses, and causal Mendelian-randomization evidence strong → we say so · thin → we say that too. Where the research is settled, we prioritize it. Where it's early, we name it plainly and don't anchor there.
Every number below traces to a named, dated, peer-reviewed source. We'd rather under-report than overstate. Coaching, not medical care — and "immortality" is the long view, not a claim on your lifespan.
How we weigh evidence
The tier is part of the reading.
Not all evidence reads the same. Every recommendation on this page carries a tier, so the rock-solid is visibly separated from the still-emerging. Most coaches say "research-backed." We show which research, and how strong — including when a popular intervention (NMN, cold plunge, fasting) is unproven. We prioritize Tier 1, and we never promote a recommendation above the tier its evidence supports.
Well-supported
Multiple cohort studies, smaller but consistent meta-analyses, mechanistic plausibility. Worth following — less certain than Tier 1, and labeled that way.
Emerging
Limited cohorts, conflicting findings, or a strong mechanism with little human-outcome data. Worth tracking; weak commitment. Reported as emerging.
Insufficient
Heavily promoted online, weak human evidence. We do not anchor here — and we name what lives here, so the hype is easy to spot.
No single number
There is no composite "longevity score." The components are weakly validated once collapsed into one figure, so we show them separately and treat that as a feature, not a gap. You read the parts.
Read what this is. Ilfirin reads the wearable and lab panel you already own — no sensor, no blood draw, no clinician of ours. It runs on your own computer, reasoned by Claude (a frontier AI model) through the Claude plan you already pay for. Coaching, not medical care. "Immortality" is the name's long view, not a promise on your lifespan.
The pentad · five domains
The whole picture, in five domains.
Movement is necessary, not sufficient. The domains below are ordered by the size of their mortality and healthspan signal in the strongest available evidence — and each marker carries its source and its tier.
1 · Movement
2 · Recovery
3 · Fuel
4 · Biology
5 · Environment
Domain 1 Movement
The largest signal — and mostly read for you.
Cardiorespiratory fitness, muscular strength, daily activity, balance. The highest-leverage domain in the strongest evidence, and most of it is captured automatically by the wearable you already own.
| Signal | The reading | Target | Evidence |
|---|---|---|---|
| VO₂max cardiorespiratory fitness |
Beats smoking, hypertension, and diabetes head-to-head as a mortality predictor, with no observed upper limit — the curve never flattens. Fitter reads lower-risk, even past elite-for-age. | Anchor on your FRIEND age/sex percentile, not a raw number — above the 75th, one decile higher year over year. Train polarized: mostly easy Zone 2 plus 1–2 hard interval sessions per week. | Mandsager 2018, JAMA Netw Open (n=122,007); Kokkinos 2022, JACC (n=750,302); Kaminsky 2017, Mayo Clin Proc (FRIEND percentiles). Tier 1 |
| Grip strength muscular strength |
A separate, continuous signal, independent of cardio fitness. The dose-response is a smooth gradient, not a cliff: each 5 kg less grip tracks ~16% higher all-cause mortality across 17 countries. | Working thresholds — men ≥40 kg, women ≥27 kg by middle age, held or improving year over year. Hand-to-hand asymmetry under ~5 kg. | Leong 2015, Lancet (PURE, n=142,861) for the dose-response; EWGSOP2 2019, Age Ageing for the clinical cut-offs. Tier 1 |
| Resistance dose strength dose-response |
Strength training cuts all-cause, cardiovascular, and cancer mortality on a J-shaped curve. Maximum benefit lands at a modest 30–60 min/week; past ~60 min it plateaus. More is not categorically better. | Two ~30-minute sessions per week, compound movements, progressive overload. | Momma 2022, Br J Sports Med (16-cohort meta); Shailendra 2022, Am J Prev Med. Tier 1 |
| Daily steps daily activity |
Mortality drops steeply with activity, then largely flattens by ~7,000–9,000 steps. The famous 10,000 isn't magic — it adds little beyond the sweet spot. | 7,000–10,000 steps on most days; cadence ≥100 steps/min during sustained walks. | Paluch 2022, Lancet Public Health (15-cohort meta); Ding 2025, Lancet Public Health; Brown 2020 (cadence). Tier 1 |
| Balance & floor-transfer mobility & balance |
Less direct mortality data than the rest, but simple field tests of balance and floor-transfer predict it — most relevant past 50, where falls drive major morbidity. | Single-leg balance ≥30 s; sit-rise test ≥8 of 11; a few minutes of mobility most days. | Brito 2014, Eur J Prev Cardiol (sit-rise); Araujo 2022, Br J Sports Med (10-second balance). Tier 2 |
Domain 2 Recovery
Sleep, autonomics, and how well you bounce back.
Sleep duration and consistency, autonomic regulation, recovery. Mostly read by the wearable — plus one self-rated check-in that, surprisingly, often beats the sensors.
| Signal | The reading | Target | Evidence |
|---|---|---|---|
| Sleep duration & regularity sleep & regularity |
Mortality is lowest near ~7 hours; both short and long sleep climb from there. Day-to-day regularity is the stronger predictor in recent data — the top quintile of sleep-regularity read 20–48% lower mortality even after adjusting for duration. | 7–8 hours most nights; hold bedtime and wake time within ~30–60 min day to day. | Yin 2017, J Am Heart Assoc (duration); Windred 2024, Sleep (UK Biobank, n=60,977 — regularity). Tier 1 |
| HRR & HRV autonomic recovery |
Both predict mortality beyond resting heart rate — and both are noisy. We render HRV as a rolling z-score against your own baseline (single-day numbers invite panic), and flag when a workout's recovery profile makes the HRR target invalid rather than scoring it wrong. | HRR <12 bpm at one minute post-effort is the high-risk threshold; ≥20 bpm is an athletic target (passive cool-down only). Resting heart rate low and stable; sustained drift is an early illness/overreach flag. | Cole 1999, NEJM (HRR threshold <12 bpm); Plews 2013, Sports Med (HRV noise caveat). Tier 1 |
| Subjective readiness how you feel |
A counterintuitive reading worth honoring: a simple morning rating of sleep, soreness, and energy is more sensitive to training-load change than HRV or submaximal heart rate. The ~30-second check-in is signal, not formality. | Daily 1–5 ratings; two low days in a row is flagged as a counterweight to the wearable. | Saw 2016, Br J Sports Med (56-study meta, n>9,000 — subjective effect sizes 2–3× larger). Tier 1 |
| Sleep-apnea screening deadly and treatable |
Untreated moderate-to-severe obstructive sleep apnea independently raises all-cause and cardiovascular mortality — and it's fixable. We screen on the symptom red-flags; where a wearable surfaces an overnight oxygen-variation signal, we'd treat it as a screening flag, never a diagnosis. | Loud habitual snoring, witnessed pauses, daytime sleepiness, or resistant high blood pressure → get a sleep study. screening Tier 1 · CPAP-for-prevention Tier 2 | Young 2008, Sleep (Wisconsin, n=1,522); Marshall 2008, Sleep (Busselton). Tier 1 |
Domain 3 Fuel
What you actually eat, against the evidence.
Diet quality, protein adequacy, fiber, ultra-processed share, alcohol. We track the pattern and show the source — no moralizing. Pattern beats perfection.
| Signal | The reading | Target | Evidence |
|---|---|---|---|
| Mediterranean pattern dietary pattern |
One of the few dietary patterns tested in a hard-outcome randomized trial — it cut major cardiovascular events ~30% in a high-risk population. | Vegetables, legumes, whole grains, nuts at most meals; fish 2+×/week; olive oil as the primary fat; little red and processed meat. | Estruch 2013/2018, NEJM (PREDIMED, n=7,447 RCT); Trichopoulou 2003, NEJM. Tier 1 |
| Fiber & UPF share fiber · ultra-processed |
Two of the strongest single dietary signals. Each extra 8 g/day of fiber tracks ~19% lower all-cause mortality; higher ultra-processed intake tracks ~21% higher — and a controlled inpatient trial showed UPF drives spontaneous overeating. | Fiber ≥30 g/day (men), ≥25 g/day (women); keep obviously ultra-processed food under ~20% of calories. | Reynolds 2019, Lancet (WHO umbrella review); Lane 2024, BMJ (45 meta-analyses, ~10M); Hall 2019, Cell Metab (inpatient RCT). Tier 1 |
| Protein a range, not a number |
The right target depends on goal state, so we hold a range rather than a dogma. Per-meal dose matters as much as the daily total. | Lean, active, at maintenance: 1.4–1.6 g/kg/day. In a deficit: up to ~2.2 g/kg. Aim ≥30 g per meal across 3–4 meals. | Morton 2018, Br J Sports Med (plateau ~1.6 g/kg); Helms 2014 (deficit); Moore 2015 (per-meal dose). Tier 2 |
| Alcohol exposure |
The "moderate drinking is protective" story has largely been retracted — the old benefit was an artifact of sick people quitting. The level that minimizes health loss is zero. Below ~25 g/day the data are methodologically contested; the load-bearing harm, for cancer, sits well above that and is causal. | Less is better; there is no minimum safe dose. We report the literature and track your actual intake — no judgment. | Griswold 2018, Lancet (GBD, 592 studies); Zhao 2023, JAMA Netw Open; Stockwell 2024 (abstainer-bias correction). Tier 1 |
| Glucose regulation TyG anchor |
Insulin resistance shows in the data a decade before a diabetes diagnosis. Our primary anchor is the TyG index — computable from any standard lipid + glucose panel, and more reproducible than fasting insulin. | HbA1c <5.7% (ideally <5.4%); fasting glucose <100 mg/dL; TyG index <4.49. | Simental-Mendía 2008; Sánchez-Íñigo 2016, Eur J Clin Invest (n=5,014); Ding 2021, Cardiovasc Diabetol (meta n=5.4M). Tier 1 |
Domain 4 Biology
Your bloodwork, read and computed.
Upload an annual physical or longevity panel. Ilfirin reads it, compares to evidence-based targets, and computes the markers most labs never surface for you.
| Signal | The reading | Target | Evidence |
|---|---|---|---|
| ApoB cardiovascular headline |
By Mendelian randomization, ApoB-bearing particles cause atherosclerotic heart disease; cumulative lifetime exposure is what counts. A more accurate marker than LDL-C, because every atherogenic particle counts roughly equally. | <90 mg/dL at average risk; <80 if elevated risk; <65 with existing disease. Plus Lp(a) <50 nmol/L (one-time) and hsCRP <1.0 mg/L. A one-time coronary-calcium scan at 45–55 sharpens the target. | Ference 2017, Eur Heart J (EAS consensus, MR); Sniderman 2019, JAMA Cardiol; Whelton 2024, JACC: Advances (CAC). Tier 1 |
| Blood pressure 120–139 is actionable |
Modern guidelines stopped treating anything under 130/80 as fine. The 2024 European guideline added an "elevated" band (120–139 / 70–89) — lifestyle change for everyone in it, treatment if it stays up in higher-risk people. | <120/80 is the anchor; 120–139 is a zone to actively lower, not tolerate. We also track pulse pressure and visit-to-visit variability. | McEvoy 2024, Eur Heart J 45:3912–4018 (ESC); Stevens 2016, BMJ + Wang 2024 (variability). Tier 1 |
| Computed markers free, from CBC + lipids |
From a standard CBC + lipid panel we derive several validated inflammation and lipid-risk markers at no extra cost — signal otherwise buried in a printout you'd file away. | NLR <2.0, RDW <13.5%, AIP <0.11, plus NHR — all computed from labs you already have. | Song 2021, Atherosclerosis (NLR); Hong 2019, Sci Rep (RDW, n=27,063); Tan 2024, Lipids Health Dis (AIP). Tier 2 |
| PhenoAge biological-age trend |
A biological-age trend, not a precise number. PhenoAge estimates biological age from ten standard lab values and predicts mortality well — but it moves with a passing illness or a hard training block, so a single reading can mislead by years. We show it as a glanceable trend, never a decision hung on one figure. No $300 epigenetic clock required. | PhenoAge ≤ your chronological age; interpret only with ≥2 measurements ≥3 months apart. | Liu/Levine 2018, PLOS Med (NHANES, HR 1.09/yr acceleration) — Tier 2, not a number to act on alone. Tier 2 |
Domain 5 Environment
The conditions you live in.
Social connection, air, light, noise, purpose, the senses. The most-overlooked domain — and several of its anchors rival the physical ones in size.
| Signal | The reading | Target | Evidence |
|---|---|---|---|
| Social connection loneliness rivals smoking |
One of the strongest non-physical signals there is — the mortality impact of social isolation is comparable to smoking ~15 cigarettes a day. We use the validated 3-item UCLA scale, and you self-report; we don't surveil your social life. | UCLA-3 loneliness score ≤4; aim for several hours of in-person time each week with people outside your household. | Holt-Lunstad 2015, Perspect Psychol Sci (70 studies, 3.4M); Wang 2023, Nat Hum Behav (90 cohorts, 2.2M). Tier 1 |
| Air quality no safe threshold |
Fine-particle (PM2.5) exposure raises mortality on a straight line all the way down — no clean cutoff below which it's "safe," just lower-is-better. Indoor air dominates total dose for most adults. | Annual PM2.5 toward the WHO 5 µg/m³ guideline / EPA 9 µg/m³ floor; in radon-prone areas, test once. | Burnett 2018, PNAS (GEMM, 41 cohorts); EPA 2024 NAAQS; WHO 2021 AQG; Darby 2005, BMJ (radon). Tier 1 |
| Light · circadian · noise cheap to address |
Bright nights and dark days each independently predict higher mortality in large actigraphy data; chronic environmental noise is causally linked to cardiovascular death. Cheap to address, easy to ignore. | <1 lux in the bedroom while sleeping; ≥30 min outdoor daylight, ideally early; keep nighttime bedroom noise low. | Windred 2024 & Burns 2023 (UK Biobank, light); Mason 2022, PNAS; Münzel 2021, Eur Heart J (noise). Tier 1 |
| Purpose · senses · screening an age-appropriate calendar |
Self-rated purpose independently predicts lower mortality. Untreated hearing and vision loss track cognitive decline. And some screenings cut mortality directly through early detection — so we keep an age-appropriate calendar. | Purpose ≥5/7 (watch sustained drops); hearing test at 50+, eye exam every 1–2 years at 40+; colonoscopy at 45, plus age/sex-appropriate cancer screens. | Cohen 2016, Psychosom Med (purpose, n=136,265); Loughrey 2018, JAMA Otolaryngol (sensory); USPSTF screening recommendations; Livingston 2024, Lancet (~45% of dementia modifiable). Tier 1 |
The honest part
What we don't promote.
These get heavy airtime in longevity-influencer space, but the human-outcome evidence isn't there yet. Saying so is part of the product. If you enjoy any of them and they're safe, that's your call — we just won't dress them up as foundational, because we have nothing to sell you.
Cold plunges
A recent meta-analysis found a brief stress dip and a short-lived quality-of-life bump that was gone by 90 days — plus a small rise in acute inflammation right after. No mortality data. The marketing has outrun the evidence.
Cain 2025, PLOS One (11 RCTs, n=3,177).
Time-restricted eating / IF
Mechanistically plausible, but the human data are mixed — a controlled trial found no body-composition advantage over normal eating when calories were matched. A fine strategy if it suits you; not a pillar.
Lowe 2020, JAMA Intern Med.
Rapamycin · metformin · NMN/NR · senolytics
Real model-organism and mechanistic data — but no human mortality or healthspan-RCT outcomes for any of them. Rapamycin's recent human trial missed its primary endpoint; metformin may blunt exercise gains in non-diabetics; NAD⁺ precursors raise NAD⁺ without a demonstrated functional benefit. We'll revisit each as outcomes land.
PEARL 2024/25 (rapamycin); Konopka 2019, Aging Cell (metformin); 2025 NAD⁺-precursor meta-analysis.
Sauna
Strong association with lower mortality — but from a single Finnish, single-sex, sauna-normalized cohort, with limited replication elsewhere. Worth including if you have access and enjoy it; not foundational.
Laukkanen 2015, JAMA Intern Med (Kuopio, n=2,315 men).
Right for some — not all
GLP-1 drugs (Ozempic / Wegovy class)
Tier 1 cardiovascular-outcome evidence — but only in the population it was tested in: people who are overweight or obese with established or high cardiovascular risk, where it cut major cardiac events ~20%. There is no evidence, and no indication, for lean, metabolically healthy people taking it for "longevity." And because a third or more of the weight lost is lean mass, anyone on one should treat protein and resistance training as more important, not less.
Lincoff 2023, NEJM (SELECT, n=17,604); Wilding 2021, NEJM (lean-mass loss 45.2%).
Solid for one thing, unproven for another
Creatine monohydrate
The cheapest, most-studied, safest training adjunct — Tier 1–2 for strength and lean mass, working through the proven strength→longevity pathway (not as a longevity drug in itself; there's no direct mortality data). The popular cognitive claim is Tier 3: the signal shows up mostly in older, sleep-deprived, or stressed people, not healthy young adults. We won't promote it above its tier.
Chilibeck 2017 (strength/lean mass); Prokopidis 2023 & EFSA 2024 (cognition — unproven in healthy adults).
Early, invite-only beta
This is the standard we coach to.
A longevity coach that cites every claim, names its evidence tiers, and says plainly when something is hype — that's the whole idea. It runs on your own computer, reasoned by Claude through the Claude plan you already pay for, and your records stay in files you own.
An early, invite-based beta — not a finished product, and not pretending to be. Coaching, not medical care. Privacy by architecture — no Ilfirin cloud, account, or database. Your records stay in files you own; during a chat, the readings Claude needs pass through Anthropic. Reads today: Fitbit · Pixel · Withings · Garmin (community) · Oura · Whoop · Apple Watch (via the free Health Auto Export app). Consumer CGM patches (Stelo, Lingo) aren't supported yet — a Dexcom/Libre bridge can push glucose in.