The science behind the panel.

Ask someone "do you sleep well?" and you get the answer they want to believe. Ask them "what % of people your age will say they sleep well?" and something interesting happens: to guess the crowd correctly, they have to consult what they actually know, including the uncomfortable parts. This is the core of our Play deck. Every card asks two things: your answer, and your prediction of how the panel will answer. The mechanism descends from a well-studied family of elicitation methods, Prelec's Bayesian Truth Serum and the "surprisingly popular" algorithm (Nature, 2017), which reward respondents for information rather than conformity. Three practical effects we see in the data: 1. Reduced social-desirability bias. When the bonus rides on reading the crowd, admitting "most days I crash after lunch" costs nothing and predicting honestly pays. 2. A built-in quality signal. Respondents whose predictions track reality (our Crowd Reader and Crowd Master badges) are demonstrably attentive. Their answers carry more signal, and clinicians can weight for it. 3. Engagement without inflation. The average member's accuracy converges around 60-75%. That gap between confidence and reality is itself a finding, and it keeps the game genuinely hard. For partner clinics the takeaway is simple: a panel that plays the crowd-guess game is a panel that answers carefully. The bonus structure is not gamification garnish; it is the data-quality mechanism.

Every number a partner or API client sees on saviorofhealth passes one gate: no slice is ever shown unless at least 5 distinct members are behind it. This is k-anonymity with k=5, enforced in one shared code path rather than per feature. What that means concretely: • A clinic sees "34 people answered, 41% chose 'Not interested'" and never who. • Answer-group health patterns (say, the average sleep of people who answered "Most days") only render when the group has 5+ members. Smaller groups display as locked. • Age-cohort breakdowns suppress any cohort under 5. Region reporting uses macro-regions (East Asia, Europe…) precisely so small-town members can't be singled out. • The AI analyst agent that partners chat with receives only these suppressed aggregates as context. It cannot answer questions about individuals because it never sees them. Members' raw journals, the water, sleep, mood and meal logs behind the statistics, stay in their account. Partners buy the shape of the crowd, not the people in it. We publish this policy in our API docs and it applies uniformly: to the dashboard, the v1 Data API, CSV exports, and the AI brief. If a future feature can't pass the k=5 gate, it doesn't ship.

A survey answer on its own is thin evidence. The reason clinics commission questions on saviorofhealth is what surrounds each answer: thirty days of real health-journal context. Take a real example from our metabolic panel. When respondents were split by their answer to a post-lunch energy question, the group answering "rarely" averaged 7.2 hours of sleep; the "most days" group averaged 6.9 and logged less weekly exercise. None of that was asked in the survey. It came from the health journals members already keep inside the app. This is the loop in full: 1. A clinician writes questions at /partner/new (duration, options, target reward) and they go live after a same-week review. 2. Members meet the questions on the Quests page, guided one by one through a conversation with an AI interviewer, and answer for Heal Points. 3. Results stream into the partner dashboard: distributions, age cohorts, daily trend, and the health-pattern table showing how each answer group actually sleeps, moves, hydrates and feels. 4. An analyst agent, grounded in exactly those aggregates, answers follow-ups and proposes the next question worth commissioning. The panel gets paid, in points, for data they were already generating. The clinic gets context no one-off survey tool can produce. And the whole exchange clears through the same k=5 anonymity gate we apply everywhere. If you run a clinic or research group and want to put a question to the panel, request access at /developers.