How this works
From a reading to something you can check.
Seven steps, in order. Each one is inspectable: you can always see what a conclusion was built from, and what it left out.
01
Collect
Readings from your meter or CGM, meals, medication, movement and sleep. Import a CSV, or type one in.
02
One timeline
Everything in the order it happened, each entry carrying its source and, when it was estimated, how confident that estimate is.
03
Find patterns
A statistical model looks for repeatable associations: post-meal responses, morning glucose, what follows a walk.
04
Weigh the explanations
A pattern usually has more than one cause. Every finding carries the other reasons that could produce it, each saying what would show it up, what is missing from your record, and whether that is something you can start logging — from a reviewed list, never written on the spot.
05
Propose a safe test
Where uncertainty is worth resolving, a small experiment you can run in a week. Every proposal is checked against your care profile first, and some are refused outright. Starting one writes down what is expected of it in the same moment, and that cannot be edited afterwards.
06
Measure the result
What happened, beside what was predicted, and the gap between them. Neither side can be revised once written — the database refuses it — so the record of how often this was right about you is the one thing here nobody can improve after the fact.
07
Turn it into evidence
Thirty or ninety days on one page: what was found, what was tested against a prediction, what it does not account for, and what is worth raising. Yours to bring, and readable in a minute.
The rule that matters
A model finds it. Language only phrases it.
Findings are produced by a statistical model. A language model is allowed to put one into a readable sentence. It is never allowed to create one, soften one, or fill a gap where the data is thin.
That is why every finding arrives with a sample count, a confidence, and a list of what it does not account for. If those are missing, the finding is wrong by construction.
When there is not enough data, you are told there is not enough data, along with what to record to change that. It is a worse-looking answer and a more useful one.
Meals followed by a walk were associated with a 1.1 mmol/L lower rise.
- What this does not account for
- This is an observed association, not a controlled comparison
- Meals in the two groups were not matched for size or composition
- Activity intensity and duration were not accounted for
- What would sharpen it
- Alternate walking and not walking after a similar meal, and log both
- Record how long and how briskly you walked
Provenance
Measured and estimated never look the same.
A reading from a CGM and a carbohydrate figure you guessed at are not the same kind of fact, and the interface never lets them look like it. Anything the platform did not observe directly carries a confidence, and that confidence is shown next to it rather than folded into an average.
The same applies to imports. Re-importing an overlapping export skips duplicates instead of overwriting them, so history cannot quietly change under you, and the original file is kept so any import can be replayed.
Safety
What is gated, and what is refused outright.
Experiments are classified before they can be started. Anything touching medication, fasting, or a major change is gated behind clinician review. Anything in the refused list can never run: asking for one is answered, and kept as a record of what was asked and why it was declined, but nothing about it is a matter of timing.
These are database constraints, not copy. An unrecognised experiment template is treated as gated rather than allowed, so the failure direction is always the safe one.
Never, under any circumstances
- Adjust your medication, or tell you to
- Calculate an insulin dose
- Diagnose a complication
- Advise you during a hypo or a hyper
- Tell you to stop taking something
Only with a clinician
- Medication timing
- Fasting protocols
- Major diet changes
- Exercise changes if you carry cardiovascular risk
Your data
Where it lives and what leaves.
Health records are stored on infrastructure we control, not in a third-party analytics tool. Photos and imported files are kept in object storage under keys that reveal nothing about you, and reach your browser only through short-lived signed links.
Every read and write of your record is logged to an append-only trail you can inspect from your own account. That trail cannot be edited, including by us, and if you ask to be erased it is your identity that is removed from it, not the record that something happened.
Questions about any of this belong on the contact page.