Health insights that wait for evidence.

Most health apps are built to alarm you.

An elevated reading on a single night is noise. Sold back to you as a notification, it becomes anxiety — and after enough false alarms, you stop reading the ones that matter.

This isn't a matter of taste. Alert fatigue is a documented clinical harm: as the volume of alerts rises, clinicians override more of them, including the ones that were right. Ancker et al. measured it in 2017, and the finding generalizes uncomfortably well to a wrist that buzzes every morning with a verdict on your night.

The competition sells urgency. Quorum sells restraint. Every design decision below follows from that one, including the ones that make the product look less capable than it is.

The name is the thesis.

Bacteria don't act alone. Each cell releases a signal molecule and listens for how much is coming back. Below a threshold, nothing happens. Once enough cells agree — a quorum — the whole population acts at once.

It is one of biology's most elegant answers to a problem every health app has and almost none solve: how do you know when a signal is real enough to act on? A single cell can't tell. Neither can a single night's heart rate variability.

So Quorum does what the cells do. It accumulates. A finding stays quiet until it has crossed a threshold that was set before the data arrived — enough days, enough persistence, enough effect size to clear the noise. Restraint isn't the absence of an opinion. It's the discipline of earning one.

An agent that shows its work.

Axiom doesn't hand you a score. It investigates — and streams the reasoning live, so you can watch it check, connect, rule out, and occasionally give up.

An axiom is a starting truth: the thing you reason from. That's the job. Every session follows the same five-step protocol, and the interesting one is step four.

The constraints are in the code, not the marketing.

Anyone can promise an honest AI. These are the rules Axiom actually runs under — several of them enforced deterministically, outside the model, because a prompt alone isn't a guarantee.

Conventional statistics, held to their actual requirements.

There's nothing exotic here, and that's deliberate. These are well-understood, non-parametric tests chosen because they're robust to the messiness of real wearable data — and, unusually, they're only ever run when their data requirements are genuinely met.

Every sync feeds a seven-phase pipeline: reconcile what changed, rebuild activity-state intervals, derive 210+ features across cardiac, sleep, activity, nutrition, and respiratory domains, then run 12 analyses and 18 composite cross-metric patterns against baselines at 7, 30, 90, and 365 days.

What makes a finding worth your attention

Significance isn't a black box. A finding's score is a weighted composite, multiplied by data quality so that a strong signal measured badly can never outrank a modest one measured well — an idea taken from Goldsack's 2020 work on evidence for digital measures.

The largest term asks whether a change is big enough to matter clinically, not merely big enough to be statistically detectable. That distinction — the minimal clinically important difference, after Jaeschke 1989 — is why Quorum stays silent through changes that a p-value would happily call significant. The rest weighs persistence, recency on a 14-day half-life, and agreement across metrics.

Thirty good days beat eight years.

We assumed history was the moat. It isn't. Almost every analysis saturates within 30 to 180 days — after that, more history buys runtime cost without buying confidence.

What matters is recent and complete. A well-covered month tells Quorum more than a patchy decade, because the tests above care about density and persistence inside a window, not the length of the archive behind it. Annual seasonality is the one genuine exception, and it says so honestly: it will not run until there are a real 365 days.

This is a better product for it. You don't have to have worn something for years to be worth talking to. And confidence is shown the way it's actually earned — “22 of the last 30 days covered,” never a raw sample count engineered to look impressive.

Every claim traces to a quote.

Roughly two thousand papers, distilled into more than nine thousand individually cited claims: interpretation facts, study findings, device and measurement caveats, and the methodology of the app itself.

The corpus was built by agents, which raises the obvious question. The answer is a rule enforced by a validator, not by good intentions: agents transcribe papers; they never invent findings. Every claim carries a verbatim quote from its source. Missing detail lowers a confidence score rather than getting filled in.

We test our own agent for fabrication

An evaluation bank resolves cited DOIs against the live Crossref API and deliberately seeds known failures — a paper with the wrong DOI, a plausible-sounding metric that doesn't exist, an invented tool name — to confirm they get caught. Hallucination is treated as a test case with an expected result, which is the only way to know whether the rules above are actually holding.

Guarded carefully. Described accurately.

The temptation in health software is to say “bank-grade” and move on. Here is what is actually true, stated at the level of detail that lets you check it.