Essays on open science, health-state mapping, responsible AI, and what it actually takes to build research infrastructure in public.
Ambitious health-AI projects often begin with a model and end with a demo. The harder, more valuable distance is from a promising result to infrastructure a whole ecosystem can trust, reproduce, and build on. A note on why we are starting with the foundation, not the headline.
Read full article
Position, trajectory, deviation from baseline, uncertainty, and evidence traceability: the five things every coordinate carries, and why each one matters.
Read moreWhy negative results belong in the open record, and how we document them alongside the ones that did work.
Read more
When you treat disease as a label to assign, you build different infrastructure than when you treat it as a trajectory to understand. The difference runs all the way down.
Read moreWe have received 501(c)(3) recognition. A short note on what that means for the mission and what it doesn't change.
Read more
How we handle physiological signals from consumer devices, including what we do when the data is noisy, incomplete, or not validated in research settings.
Read moreFindable, accessible, interoperable, reusable: we explain how these principles are baked into the way we store and share every dataset and model artifact.
Read moreDiagnostic categories name syndromes; biologically grounded subtypes point toward the variation that actually drives them.
Read morePrediction and causal inference are different problems with different requirements; conflating them produces unreliable infrastructure.
Read moreOn-device computation is a design principle that puts the person at the center of their own health data, not a limitation to work around.
Read moreReleasing model weights is a good start; genuine openness requires training data provenance, evaluation code, model cards, and the negative results too.
Read moreA mission-locked nonprofit can pursue long time horizons and full openness that neither academic labs nor commercial companies are structured to sustain.
Read moreHealth exists on a continuum; infrastructure that treats it as a set of discrete labels loses information at every step of the science.
Read moreWhat looks normal for a population may not be normal for a specific person; meaningful change can only be detected against that person's own history.
Read moreBiology is measured in many modalities simultaneously; models that integrate them offer a richer picture of health state, but only if done carefully.
Read moreA model that performs well on average can still perform poorly for specific populations; discovering that gap before deployment is an ethical requirement.
Read moreReproducibility built into a pipeline from the start is far cheaper and more durable than reproducibility retrofitted as documentation after the fact.
Read moreStructured documentation turns a model or dataset from an opaque artifact into something the broader community can evaluate and hold accountable.
Read moreThe people whose lives are most affected by health research deserve a role in shaping it, not only in contributing data to it.
Read moreCytognosis is research infrastructure and does not provide medical diagnosis, treatment, or clinical decision-making.