Most health hacks do not survive measurement.
InVivo turns wellness folklore into small, safe, falsifiable personal experiments — then aggregates the results through federated evidence. Run a claim against your own physiology, and your anonymized result helps falsify or refine it for everyone else. A truth engine, not a recommendation engine.
every claim must pass before it can be tested
supported, falsified, inconclusive, harm — never 'proved'
post-meal walk, caffeine, wake time, magnesium, blue light
only privacy-preserving aggregates leave your phone
“Magnesium improves my sleep.”
In adults with self-reported poor sleep and no safety exclusions, does magnesium glycinate taken nightly for 14–28 days improve sleep latency, efficiency, next-day energy, or HRV compared with placebo or control?
Minimum worthwhile effect
Without a pre-specified threshold, every tiny wearable fluctuation becomes 'a result'. A claim is not supported when the federated estimate excludes the minimum worthwhile effect — or when the benefit is smaller than the burden, risk, or cost.
Stricter than 'proved'
Verdict cards never say a claim is proved. They report supported-for-you, falsified-for-you, not-supported-in-cohort, a subgroup signal, inconclusive, or a stop — each with its uncertainty.
Every claim must pass five gates before you can test it.
Most social-media health claims fail at least one. That is the point — the gates are a filter, not a formality.
Is it safe enough to test?
A contraindication and medication-interaction screen runs first. Pregnancy, kidney disease, prescribed-medication changes, and high-risk protocols are excluded or routed to a clinician.
Is the outcome measurable?
The claim must map to a signal your phone, wearable, or labs can actually capture — sleep latency, post-meal glucose iAUC, home blood pressure, HRV, a validated score.
Can it be falsified in reasonable time?
Reversible claims with fast outcomes become n-of-1 crossovers. Slow biomarkers become lab-anchored before/after. Claims that can never resolve are marked research-only.
Can placebo and expectation be reduced?
Randomized A/B/B/A periods, washouts, and — where possible — blinding separate a real effect from the wish for one. A breathing practice is compared to quiet rest, not to nothing.
Can results aggregate without exposing data?
Your phone computes the effect locally and contributes only a masked summary through secure aggregation. No individual result is ever seen in the clear.
InVivo does not say “try this hack.” It says: this claim is popular, the evidence is uncertain, and you can run a small experiment to see whether it works for you.
Claim → Protocol → Personal experiment → Federated evidence → Verdict.
Each claim becomes a versioned object carrying its population, exclusions, endpoints, minimum worthwhile effect, design, washout, safety checks, evidence prior, and verdict rules.
Claim
Exact wording, mechanism, population, exclusions, endpoints, minimum worthwhile effect.
Protocol
Design, duration, washout, safety checks, evidence prior, verdict rules — a versioned object.
Personal experiment
Baseline → randomized A/B periods → washout, with adherence and outcomes logged on-device.
Federated evidence
Effect, standard error, adherence, and safety events aggregated through secure aggregation.
Verdict
A public verdict card with uncertainty: supported, not supported, subgroup signal, or stop.
The number that stops noise becoming a result.
Each endpoint declares the smallest effect worth acting on. Anything smaller is treated as no effect, regardless of how confident the statistics look.
Five designs, matched to how a claim can actually be falsified.
The design follows the biology: reversible and fast becomes a crossover; slow biomarkers become lab-anchored; intermittent conditions become event-triggered; the rest stays observational.
Personal n-of-1 crossover
Reversible claims with fast outcomes: caffeine cutoff, magnesium, post-meal walking, breathwork. Multiple randomized A/B periods within one person.
Micro-RCT at population scale
Low-risk nudges: morning light, post-meal-walk prompt, earlier dinner. Eligible users randomized; outcomes collected locally, effects federated.
Lab-anchored before/after
Slower biomarkers: vitamin D correction, omega-3 and triglycerides, creatine and strength. Baseline labs, 8–12 weeks, matched comparator, safety monitoring.
Event-triggered trials
Intermittent conditions: zinc at first cold symptom, electrolytes during long runs, curcumin during knee-pain periods. Randomized at the trigger.
Observational-only
Too risky, slow, or confounded for self-experiment — sauna and heart attacks, cold plunges and longevity. Tracked as cohorts or avoided, never recommended as a hack.
Seven verdicts. None of them is the word 'proved'.
The output uses a stricter evidence vocabulary than wellness apps, so a personal result and a cohort result are never confused with universal truth.
Supported for me
Personal effect exceeded the threshold with adequate adherence.
Falsified for me
Personal experiment showed no worthwhile effect for this user.
Not supported in cohort
Federated effect fell below the minimum worthwhile effect.
Supported in subgroup
A pre-specified subgroup showed a meaningful effect.
Inconclusive
Too little adherence, data, or power to decide.
Potential harm / stop
Adverse effects, worsening outcome, or a safety signal.
Research-only
Interesting, but not ready for any recommendation.
Blue-light glasses improve sleep.
Not supported for general sleep improvement.
No meaningful objective sleep benefit in the overall InVivo cohort. Possible subgroup signal in users with insomnia symptoms and high evening screen exposure.
Recommendation
Do not prioritize over caffeine timing, regular wake time, and light exposure.
10-minute post-meal walking lowers glucose spikes.
Supported for post-meal glucose control.
Meaningful reduction in post-meal glucose excursion, strongest in users with elevated baseline glucose variability.
Recommendation
Offer as a first-line metabolic micro-protocol before supplement experiments.
The first five claims are measurable, reversible, and likely to break.
They are chosen to produce both supported and falsified results quickly — the fastest way to show that measurement, not marketing, decides.
A 10-minute walk after meals flattens glucose spikes.
Meal-level crossover: walk vs usual routine
Post-meal iAUC, peak glucose, time above threshold
No caffeine after lunch improves sleep.
14–28 day randomized crossover vs usual caffeine
Sleep latency, awakenings, efficiency, next-day energy
A consistent wake time beats chasing 8 hours.
4-week sleep-regularity protocol vs usual sleep
Sleep regularity index, efficiency, fatigue, mood
Magnesium improves sleep.
Blinded n-of-1 crossover, 14 days per arm
Sleep latency, efficiency, insomnia score
Blue-light glasses improve sleep.
Semi-blinded crossover: blocking vs clear lenses
Sleep latency, efficiency, subjective sleep
Your phone does the science. Only privacy-preserving aggregates leave it.
Federated analytics first, federated learning second. The phone computes the individual effect estimate, standard error, adherence, missingness, baseline risk, safety events, and burden — the server receives only secure-aggregated summaries.
Personal evidence
- Intervention adherence
- Outcome time series
- Personal effect + uncertainty
- Adverse effects and burden
- “Was it worth it?” rating
Federated cohort evidence
- Mean treatment effect + heterogeneity
- Subgroup and adherence-adjusted effects
- Adverse-event and dropout rates
- Small-cell suppression
- Survivorship and publication-bias checks
Clinical-grade evidence
- Pre-registered protocol
- Governance review where needed
- CONSORT-style reporting
- External validation
- Clinician review + public verdict card
Federated learning does not automatically solve privacy — model updates can leak. InVivo pairs it with secure aggregation, differential privacy where appropriate, subgroup suppression for small cells, and per-claim model cards, so no individual contribution is ever seen in the clear and raw signals never leave the phone.
Not a playground for unsafe experimentation.
Some claims are excluded from general consumer experiments entirely, or restricted to clinician-screened subgroups. The boundary is part of the product.
Patient-led experiments → governed evidence
ProvidEHR turns consented patient experiments into governed prevention evidence: source-grounded, federated, and linked to clinical outcomes — without centralizing raw personal data.
Tested across thousands of personal experiments
InVivo can say who a claim helped, who it did not, and when the burden or risk was not worth it — preserving privacy throughout. That is stronger than wellness content, influencer claims, or generic AI advice.