A fact needs a history.
We plan to preserve where information came from, when it happened and who checked it.
Our approach
Trust begins
with clarity.
What’s known. What’s missing. Where it came from. Those distinctions belong in the experience—not just the fine print.
ILLUSTRATIVE / ROSA FERREIRA
Where she points matters.
● Hypothetical patient mark · 1 Sep
right lateral torso. Separate input—not inferred from “right side.”
Unknown should stay unknown.
A missing result is not a normal result. An AI proposal is not a verified fact.
Learning starts with listening.
Corrections should be attributable, reviewable and visible over time.
INFORMATION, WITH ITS LIMITS VISIBLE
The gaps belong
in the picture.
These illustrative records show the design commitment: expose the information we don’t have, and the sources that disagree. A chart should not hide either.
ILLUSTRATIVE / ROSA FERREIRA
■ Ferritin · ng/mLMissing is not normal.
2022
24 ng/mLExample laboratory result
RESULT DATE · 6 Feb2023
18 ng/mLExample laboratory result
RESULT DATE · 19 Jan2024
No resultDraw recorded; result unavailable
DRAW DATE · 14 Mar2025
No resultNo result; testing status unknown
PERIOD SHOWN · 17 Feb2026
11 ng/mLExample laboratory result
RESULT DATE · 20 AugILLUSTRATIVE / ROSA FERREIRA
Conflict unresolvedOne medication. Three accounts.
LevothyroxineDocumented doses, not dosing advice
| Source | Recorded dose | What we know |
|---|---|---|
| ■ Example · Brooklyn primary care31 Jul 2026 | 112 mcg | Listed dose · unverified |
| ● Your entry1 Sep 2026 | 137 mcg | Dose you gave · unverified |
| ■ Example · Providence urgent care8 Jan 2026 | 112 mcg | Stale entry · current use unknown |
Different doses. No silent decision.
The example preserves the discrepancy. It does not select a dose, recommend a change or treat three records as three prescriptions.
● Your entries■ From a clinic, lab or pharmacy▧ Missing information
Behind the approach.
Bounded rules, not model-made calculations
Medication checks would use versioned rules and supported inputs. Each result would link to its rule and evidence. A language model could explain a checked result; it would not calculate a dose or interaction severity.
Deterministic means reproducible, not automatically correct. Rules still need source and licensing review, expert validation, formulation and unit checks, and explicit abstention when inputs are unsupported.
Human authority, clearly attributed
AI output belongs in a proposal queue, separate from accepted facts. A human transition needs an actor, a time and a verification state. Confirming what you reported is different from a clinical determination.
A future professional service would require contracted clinicians and a clear chain of responsibility. That service does not exist today.
Test the whole experience
Our evaluation plan includes source preservation, extraction, retrieval, rules, explanations and user understanding. Timeouts, errors and abstentions count—not just successful answers.
We want to measure whether relevant history makes a returning experience more useful. Severe known failures must block release, not disappear into an average.
Corrections without uncontrolled learning
A correction should record what changed, where, why and by whom. An override is not automatically proof the AI was wrong.
Any future training use would require appropriate permissions, quality review, dataset preparation, evaluation and an approved release. Editing a record would not retrain a model in real time.
A line we won’t cross
Your health story.
Never our product to sell.
We will never sell patient data. This website has no advertising pixels, session replay, health-intake form or patient account.
Read our privacy noticeThis is only the beginning