◌
Check the signal
Coverage, missingness, and wearable noise are assessed before inference. Insufficient data leads to abstention.
DATA QUALITY GATEREHEALTH AI / 睿禾健康
Health trends,
with evidence attached.
A transparent research system for turning wearable signals into grounded, readable trend reports. Every conclusion has a quality gate, a traceable fact, and a clear limit.
TREND INTELLIGENCESYNTHETIC SAMPLE
DATA QUALITY91%Example coverage
RESTING HEART RATE7-DAY VIEW
MONTUEWEDTHUFRISATSUN
Every sentence grounded in validated facts
✦Evidence first
A RESEARCH SYSTEM BUILT FOR TRANSPARENCY
01 / THE APPROACH
Designed to explain
how it knows.
Each stage has a distinct job. Uncertain data is allowed to stop the chain before it becomes a confident-sounding report.
⌁
Build the facts
The system compares validated metrics with a personal baseline and keeps the supporting evidence attached.
STRUCTURED INFERENCE✦
Write within bounds
Report text is constrained by the fact packet, checked for unsupported claims, and labeled as AI-generated content.
FACT-CONSTRAINED SYNTHESIS02 / THE EVIDENCE
The benchmark
includes the misses.
Our public wearable PPG experiment uses subject-isolated evaluation. The current candidate has not met the frozen release threshold. That decision is published with the scores.
Inspect full benchmarkPPG RESEARCH CANDIDATE / V0.5RELEASE BLOCKED
MEAN ABSOLUTE ERROR19.70 bpm
Gate ≤ 10 bpmP90 ABSOLUTE ERROR56.06 bpm
Gate ≤ 15 bpm8 subjects · 18 evaluable recordsNOT FOR LIVE HEALTH CONCLUSIONS
03 / SEE THE FORMAT
A report you can
read critically.
This fixed illustration uses synthetic values. It shows the intended report structure, not a validated health assessment.
✦SYNTHETIC EXAMPLE
TREND SNAPSHOT
A clearer viewof the week.
SAMPLE REPORT
SIMULATED DATAYour weekly trend
DATA COVERAGE91%
RESTING HEART RATE72 bpm
PERSONAL BASELINE66 bpm
What the numbers show
In this simulated week, the example resting heart rate is above the example baseline. Activity, sleep, and measurement quality should be considered together before interpreting this change.
ⓘIllustrative only. No diagnosis, treatment advice, or live personal data is used.
04 / BUILT-IN LIMITS
Knowing when
to stop.
01
Abstain on weak data
Insufficient coverage blocks a risk conclusion.
02
Keep generation grounded
Unsupported medical claims trigger a safe fallback.
03
Block unqualified releases
Models that miss quality or sample-size gates cannot be promoted.
OPEN RESEARCH / HONEST LIMITS
Follow the work.
Question the evidence.
Explore the source, methods, and complete results in the public repository.
View on GitHub