Which population is represented?
The result names the official table, geography, population group, and publication year behind the starting estimate.
Evidence and accuracy
How accurate are life expectancy calculators?
A longevity estimate is only useful if you can see how it was built and where it breaks down. Til100 publishes that, for free, inside the app — and describes the standard here.
Current path is personal context matched to an official population baseline.
Til100 currently uses confirmed age, region, and applicable sex-at-birth to match an official life table. This makes the starting context personal, but it is not an exact end date or a complete personal prediction. Stronger path stays unnumbered until confirmed clinical inputs and the matching model pass external and subgroup validation.
Why no single number
No single percentage can describe prediction accuracy.
A population life table describes outcomes for groups, not what will happen to one person. Til100 publishes its source, population, year, method, uncertainty, and limitations together instead of turning them into a personal accuracy score.
What does the range mean?
The range describes a conditional population distribution. It is not a promise, a personal confidence interval, or the headline.
What does health data change?
In this version, chosen health data supports trends and actions. It does not change the population starting estimate.
What the model card contains
See what the model uses, where it works, and where it does not.
- Who the model is for
- Who the model was built for, which markets it has been validated in, and the decisions it is not suitable for.
- What data it uses
- Every factor the current version uses, how each exposure is defined, and which candidate factors are collected but deliberately excluded from scoring.
- How it was built and tested
- The datasets behind the model, their size and time window, and how the validation cohort differs from the development one.
- How well it performs
- Calibration, discrimination, and Brier score, each with confidence intervals and tied to a dataset, horizon, model version, and review status.
- How results differ across groups
- Where performance differs across groups, reported even when the difference is unflattering.
- What it cannot do
- What the model cannot see, where it is likely to be wrong, and what would have to change for that to improve.
- What changed between versions
- Every model and evidence version, what moved, and when.
22 profile data points · public basis
Why Til100 asks for 22 health details.
Each point has a declared job and a traceable research basis. The questions are grouped into a short flow; they are not 22 separate screens. Chinese population evidence is shown first when it covers the construct, with international cohorts used where needed.
Research relevance is not model admission.
Today, only age, region, and applicable sex at birth select Current Path. The other 19 points improve profile coverage, action context, limitation checks, or readiness for a separately approved regional model. They do not add or subtract lifespan years in the Early model.
Official starting point
- AgeSelects the official life-table row.
- Country or regionSelects the closest supported population source.
- Applicable sex at birthSelects a sex-specific table only when the official source provides one.
Body context
- HeightImproves body-context coverage and regional-model readiness.
- WeightImproves body-context coverage and regional-model readiness.
- WaistPrepares Chinese cardiovascular model readiness.
Research basis
China-PAR cardiovascular risk equations ↗Everyday function
- Self-rated healthAdds questionnaire health context and action priority.
- Usual walking paceAdds functional-health context; it does not diagnose fitness.
- Movement limitationAdds functional limitation and result-boundary context.
Smoking and alcohol
- Smoking statusCan prioritize cessation support; it does not change Early headline years.
- Alcohol frequencyAdds lifestyle context without treating low intake as protective.
Diagnosed conditions
- High blood pressureChecks model fit and the need for confirmed blood-pressure data.
- DiabetesChecks model fit and the need for confirmed glucose or HbA1c.
- Heart disease or strokeFlags clinical-history context and model limitations.
- CancerFlags clinical-history context and model limitations.
- Kidney diseaseFlags clinical-history context and model limitations.
- Chronic lung diseaseFlags clinical-history context and model limitations.
Treatment and wider health context
- Blood-pressure medicineDistinguishes treated status for future validated model eligibility.
- Cholesterol-lowering medicinePrevents treated and untreated measurements being interpreted as equivalent.
- Long-term illness or disabilityAdds limitation and transportability context.
- Hospital stay or serious injury in the past yearFlags recent instability that a broad population result cannot see.
- Premature heart disease or stroke in a parent or siblingPrepares family-history context for regional cardiovascular modeling.
Public algorithm · daily-signal-v2
Today Signal is versioned, inspectable, and separate from lifespan.
Today Signal answers one limited question: how recent passive health patterns compare with your own baseline. It is a 0–100 product feedback index—not age, diagnosis, mortality probability, or a clinical score.
- Early and personal windows
- Early Signal needs all three recent days plus all seven preceding days for one factor. Personal Signal uses at least four of seven recent days and fourteen days from the preceding day 8–90 window.
- Inputs and product weights
- Activity 35% · sleep-duration stability 30% · resting heart rate 20% · HRV 15%. Available weights are normalized.
- Eligibility
- Early Signal needs one complete factor. Personal Signal needs at least two personally ready factors, 50% configured factor coverage, and data no more than three days stale.
- Formula
- component = clamp(50 + 50 × tanh(directional delta ÷ 0.25)); Today Signal = round(50 + coverage reliability × maturity × (weighted components − 50)). Early maturity is 0.5; personal is 1.
- Missing and duplicate data
- Missing, insufficient, or stale data produces Not enough data. On one metric-day Til100 selects one highest-coverage source and never sums duplicate device records.
- Display bands
- 80–100 Moving forward · 60–79 Holding steady · 40–59 Rebuilding · 0–39 Needs attention. These are feedback bands, not clinical or mortality thresholds.
The Today Signal score never triggers a Current Path change.
A future Path Review is decided per approved underlying factor—not from this composite score—and requires factor-specific evidence, approved model use, at least 21 valid days in 28, a material-change rule, and 90-day confirmation. Current wearable inputs remain coaching-only and are not used by the lifespan model.
Original research
Open the reports—and see the inference boundary beside each one.
- Self-regulation mechanisms in health behavior change ↗
- Supports personalized feedback and self-monitoring; it does not validate the Til100 score or a lifespan threshold.
- Just-in-Time Feedback systematic review ↗
- Supports personalized, actionable feedback; it does not prove a longevity effect.
- Daily steps and all-cause mortality meta-analysis ↗
- Supports a nonlinear population association; no single day is converted into lifespan.
- Sleep regularity and mortality ↗
- Supports sleep regularity as a candidate longitudinal signal; the association is observational.
- Resting heart rate and mortality meta-analysis ↗
- Supports population association, not a causal daily lifespan change.
- HRV and mortality systematic review ↗
- Supports HRV as a candidate signal; measurement differences keep it coaching-only here.
- Dynamic prediction with longitudinal markers ↗
- Supports validated longitudinal models, not a threshold on a display score.
- TRIPOD+AI prediction-model guidance ↗
- Defines transparency and validation expectations; it does not approve a Til100 model.
The evidence gate
A factor must pass external and subgroup validation before it can change headline years.
A health factor may enter production scoring only after its source set, exposure definition, effect shape, confounding limitations, transportability, coefficient extraction, and statistical and safety approval are recorded in the evidence registry.
- Sleep, smoking, alcohol, diet, and body composition remain candidate scoring domains until each passes that gate individually.
- Being able to chart a metric is not permission to score it. Display eligibility and model eligibility are separate decisions.
- Alcohol is not given a protective low-dose effect by default.
- An input with no approved, visible use is not collected on the chance that a future model might want it.
The role of AI
AI
AI can explain your result, but it cannot create it.
A deterministic service owns the population starting estimate and every approved health output. The AI layer reads approved structured facts and reviewed sources to explain, prioritize, and draft. It cannot diagnose, recommend medication or supplements, or turn an unconfirmed inference into a health fact.
In the app
See the exact source behind your starting estimate.
The app shows the population, geography, year, method, and limits behind the number. Connected health data is separately source-labelled in trends and actions.
Release status →