Your data, transparent.

01The Default Is Local

halfmarble works completely, fully, and powerfully without sharing any data beyond your own device. You get every feature, every dose calculation, every AI answer, without contributing a single byte to our backend. That's the default, and it is never changed without your explicit action.

If you choose to join the fight beyond your own case, you can opt in to share anonymized, aggregated patterns with researchers. Everything below explains exactly what that means.

02What Gets Captured

SteadyDoseTool knows both when you dose and when you eat. That combination is genuinely novel research material — the medical literature on levodopa absorption under real-world fasting and eating patterns is surprisingly thin. No existing study captures this at the scale we can.

When you ask a question to the AI in either SteadyDoseTool or SteadyFastingTool, the same rules apply. Your conversation text stays on your device. What leaves (if you opt in) is only the topic cluster and your feedback on how useful the answer was.

For Parkinson's patients, cognitive fluctuations matter as much as motor fluctuations. Brain fog, mood, apathy, and mental fatigue track closely with medication cycles and metabolic state — but nobody has this data at scale because clinical scales are administered in clinics once or twice a year. Halfmarble captures it continuously at home so you can see your own patterns and (optionally) contribute to research. Mental state data is subject to stricter safeguards than other categories.

Parkinson's is traditionally assessed in a clinic every 3-6 months using the MDS-UPDRS scale. Continuous home-based motor tracking is missing from the literature. SteadyDoseTool captures it two ways: focal symptom logging (track body-specific symptoms like foot dystonia, hand cramping, by location) and objective tapping tests that use the phone's touchscreen to measure motor function — validated against clinical UPDRS Part III finger tapping.

The phone test takes 10-20 seconds. Results correlate with dose timing automatically. Over time, you see objectively whether your predicted PK peaks match your actual motor peaks. No existing dataset combines PK prediction + objective motor measurement + dose timing + fasting state at scale.

Glucose Ketone Index (GKI) was developed by Thomas Seyfried (one of the researchers in our Glass Box) as a measure of the depth of therapeutic ketosis. Three of halfmarble's Foundational 60 researchers — Seyfried, Phillips, and Cooper — use GKI in clinical practice. No large-scale real-world GKI dataset for Parkinson's patients exists. SteadyFastingTool users who opt in contribute to the first one.

Halfmarble supports both blood finger-prick dual meters (Keto-Mojo GK+, Precision Xtra) and electrochemical breath ketone devices (KETOSCAN Smart, Biosense) paired with glucose meters or CGMs. Electrochemical breath acetone sensors are measurement-grade — validated against laboratory reference at R² >0.97 — not trend indicators.

03How It's Anonymized

Before any byte leaves your phone, your data goes through four stages of transformation. The raw data never leaves. Here's a concrete example for a single dose log:

k-anonymity with k≥50. A user's pattern only leaves the device after it can be combined with at least 49 others. If your pattern is too unique to be anonymized, it is held in a local queue and discarded after 30 days — never shared.

Differential privacy noise (ε=1.0) is added to all counts before transmission, so that even aggregated statistics cannot be reverse-engineered to individual users.

No IP addresses logged. Requests route through a relay that strips origin information server-side.

04We Will Never

A list of things we will not do is often more useful than a list of things we will. These are binding commitments — part of halfmarble's formal business model and corporate governance. We are legally obligated to uphold them, not just morally committed. A violation would invalidate the entire trust model this project is built on.

IRB = independent ethics board that reviews human research studies. Published findings = the research becomes public knowledge, not a pharma asset.

05You Can Always Leave

Changing your mind is a one-tap operation. Settings → My Contributions → "Delete my shared data." Your records are marked for deletion immediately, and removed on the next aggregation cycle (at most 7 days).

You can opt out of individual categories (e.g., keep sharing PK data but stop sharing AI feedback) without losing access to any app features. Opt-in is never required.

You can also request a full export of everything you have ever shared. Because we never link patterns back to your identity, this is computed using your device's local contribution log — another reason the raw data never leaves.

06Why This Matters

There is a version of this app that could quietly collect raw data, sell it to the highest bidder, and burn trust later. We are not building that version. Halfmarble's commercial engine is Real-World Evidence (RWE) — anonymized Home Usage Data that helps researchers understand how chronic disease patients actually live and treat themselves at home. That only works if it is transparent, opt-in, and defensible to both patients and regulators.

The people fighting the same fight as us are the ones who make this work. Without their explicit, informed contribution, there is no model, no data, no research. Which is why every decision here starts from the same place: the default is local, contribution is a choice, and the choice is fully informed.