InVivo
InVivo Experiments — the anecdote falsification engine

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.

Safety gates
5

every claim must pass before it can be tested

Verdict words
7

supported, falsified, inconclusive, harm — never 'proved'

First claims
5

post-meal walk, caffeine, wake time, magnesium, blue light

Raw data shared
0

only privacy-preserving aggregates leave your phone

The claim is translated, not repeated
Folklore

“Magnesium improves my sleep.”

Falsifiable question

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?

Design
28-day crossover
Primary
sleep latency
MWE
15 min faster onset
Boundary
wellness only
Why it matters

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.

The honest output

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.

Five gates

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.

The product promise

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.

Evidence architecture

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.

01

Claim

Exact wording, mechanism, population, exclusions, endpoints, minimum worthwhile effect.

02

Protocol

Design, duration, washout, safety checks, evidence prior, verdict rules — a versioned object.

03

Personal experiment

Baseline → randomized A/B periods → washout, with adherence and outcomes logged on-device.

04

Federated evidence

Effect, standard error, adherence, and safety events aggregated through secure aggregation.

05

Verdict

A public verdict card with uncertainty: supported, not supported, subgroup signal, or stop.

Minimum worthwhile effect

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.

Sleep latency≥10–15 min faster onset
Sleep efficiency≥3–5 percentage points
Resting systolic BP≥3 mmHg for low-risk changes
Post-meal glucose≥10–15% lower postprandial iAUC
Pain≥1 point on 0–10 or validated MCID
Stressmeaningful PSS/GAD-7 or daily-affect change
Adherence burdenbenefit must exceed self-reported burden
Experiment types

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.

A

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.

B

Micro-RCT at population scale

Low-risk nudges: morning light, post-meal-walk prompt, earlier dinner. Eligible users randomized; outcomes collected locally, effects federated.

C

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.

D

Event-triggered trials

Intermittent conditions: zinc at first cold symptom, electrolytes during long runs, curcumin during knee-pain periods. Randomized at the trigger.

E

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.

Verdict vocabulary

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.

Verdict card

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.

Burden
Low
Risk
Low

Recommendation
Do not prioritize over caffeine timing, regular wake time, and light exposure.

Verdict card

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.

Burden
Moderate
Risk
Low

Recommendation
Offer as a first-line metabolic micro-protocol before supplement experiments.

V1 catalogue

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.

01

A 10-minute walk after meals flattens glucose spikes.

Likely supported
Design
Meal-level crossover: walk vs usual routine
Primary endpoint
Post-meal iAUC, peak glucose, time above threshold
02

No caffeine after lunch improves sleep.

Strong candidate
Design
14–28 day randomized crossover vs usual caffeine
Primary endpoint
Sleep latency, awakenings, efficiency, next-day energy
03

A consistent wake time beats chasing 8 hours.

High-value prevention
Design
4-week sleep-regularity protocol vs usual sleep
Primary endpoint
Sleep regularity index, efficiency, fatigue, mood
04

Magnesium improves sleep.

Ideal falsification target
Design
Blinded n-of-1 crossover, 14 days per arm
Primary endpoint
Sleep latency, efficiency, insomnia score
05

Blue-light glasses improve sleep.

Likely falsified overall
Design
Semi-blinded crossover: blocking vs clear lenses
Primary endpoint
Sleep latency, efficiency, subjective sleep
Then the next five
Beetroot juice and blood pressureApple cider vinegar and post-meal glucoseCreatine and cognitive fatigueCollagen and joint painProbiotics and IBS
Federated proof

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

stored locally
  • Intervention adherence
  • Outcome time series
  • Personal effect + uncertainty
  • Adverse effects and burden
  • “Was it worth it?” rating

Federated cohort evidence

aggregated, no raw data
  • 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

promoted only when earned
  • 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.

Safety boundary

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.

Stopping or replacing any prescribed medication — never a consumer experiment.
Statins, antihypertensives, glucose-lowering drugs, anticoagulants, hormones, or thyroid medication unless clinician-directed.
Mega-dose fat-soluble vitamins, iron, iodine, or selenium.
Long or dry fasts, dehydration protocols, detoxes, and cleanses.
Mouth taping with snoring, nasal obstruction, suspected apnea, or sedative use.
Cold plunges with cardiovascular disease, arrhythmia risk, uncontrolled hypertension, or pregnancy.
Ashwagandha, berberine, and green-tea-extract weight-loss protocols carry strict exclusions and clinician routing.
For ProvidEHR

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.

The moat

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.

Join the experiments preview