Stable aggregates the target
Resting-heart-rate trends, sleep architecture, adherence rates, the slow drift of a cognitive baseline. Averages over the chaos beneath them — highly forecastable. This is where diagnostics live.
Mantis
Behavioral forecasting infrastructure
The last system without a digital twin is the human being.
Mantis builds living, predictive models of individual people — their physiology and behavior, forecast to the horizon where a decision still has to be made. Starting with medicine.
The twin: a living model assembled point by point from a person's record. Raw behavior tears away onto a strange attractor — chaotic, unforecastable — and the model pulls it back into form.
The record
A person is now measured continuously.
The average American spends roughly 40% of their waking life using a computer — nearly two decades of recorded interaction over a lifetime. Every one of those interactions is a measurement of a person's state. Fed into the right models, measurements of the state become forecasts of the trajectory.
≈ 5,000
digital interactions per person, per day
≈ 250,000
physiological measurements per day, from one smartwatch
93%
measured upper bound on the predictability of human movement
The theory
Chaos at the bottom. Stability at the top.
Human behavior is a chaotic system — small uncertainties compound until any forecast decays into noise. The atmosphere's forecast horizon is about two weeks, which is why weather prediction dies there. Our thesis: raw behavior is chaotic, but the correct coarse-grainings of it are not. The same averaging that lets climate be forecast for decades separates behavior into three regimes:
Stable routines reachable
Habit- and constraint-anchored patterns: where you sleep, your commute, your spending categories. Measured at up to 93% predictability in the Science mobility work.
Micro-chaos the long game
The next word you'll type, one impulse purchase, a momentary mood spike. Beyond today's models — and exactly where we're headed. The long-term goal is perfect prediction.
The questions worth answering today live almost entirely in the first two regimes — and that's where we start. Every model we ship pushes the horizon deeper into the third.
The company
A neutral layer, like OAuth for behavior.
Google, Apple, and Meta each hold a deep slice of the human record — but each sees only its own products, and each monetizes the data through one narrow business. A holistic model of the individual sits permanently outside every incumbent's mandate. The opportunity isn't too big for them. It's structurally incompatible with them.
Mantis plays the role OAuth plays for identity, Plaid for banking, and Stripe for payments: a broker that succeeds precisely because it doesn't compete with either side. Our pipelines ingest any stream a person grants us — device telemetry, wearable physiology, digital activity, clinical records — and our models turn that unified record into a living, predictive twin. Companies then request scoped, revocable access to predictions from that twin, the way an app requests access through OAuth.
The person
grants data streams, holds the keys
Mantis
unified record → predictive twin
Builders
scoped, revocable predictions
We don't sell data. We sell the connection — with the person's consent on one side and the person's interest on both.
Applications
One layer. Every product that needs a forecast of one person.
Everything below is the same primitive — a scoped, consented prediction served from the twin — pointed at a different decision.
Pre-symptomatic detection
Flag infection from wearable physiology at or before symptom onset — up to nine days before the person feels it.
Cognitive baseline monitoring
Track the slow drift of an individual's cognitive baseline and surface decline months before it would show in a clinic visit.
Decompensation forecasting
Forecast heart-failure decompensation and readmission risk inside the two-week window where intervention still changes the outcome.
Adherence prediction
Predict which patients will miss medications, and when — before the gap shows up in outcomes instead of after.
Recovery trajectories
Model expected post-operative recovery and alert the care team the moment a patient deviates from their twin's predicted curve.
Relapse early warning
Detect the physiological and behavioral signature that precedes psychiatric relapse, while there is still time to reach the person.
Trials on twins
Match patients to trials by predicted trajectory, and run counterfactual arms against digital twins instead of placebo groups.
What we haven't thought of
Any domain where a consented forecast of one person beats a population average. The layer is the product — builders decide the rest.
Evidence
The ceiling has been measured.
-
A 93% upper bound on the predictability of an individual's location, nearly invariant across millions of people.
Song et al., Science (2010) -
With ~300 Facebook Likes, a model judged personality as accurately as a person's own spouse.
Kosinski et al., PNAS (2013, 2015) -
81% of COVID-19 cases showed physiological warning signs at or before symptom onset — some nine days early — from a smartwatch alone.
Mishra et al., Nature Biomedical Engineering (2020)
The signal that you are getting sick exists in your data before it exists in your awareness. What's missing is the layer that puts that reading to work for the person being read.
Where we begin
Medicine first.
We start in healthcare because it's where individual-level prediction is worth the most, where permission to hold sensitive data is most clearly defined, and where the bar for privacy, validation, and accountability is highest. If it works here, it works.