# HealthWatch

> HealthWatch is a private, invitation-only health-research companion for people who track everything — and a health-reasoning substrate that AI agents can call. It fuses one person's labs, CGM, wearables, and FHIR records with 1,400+ curated longevity sources and PubMed, then returns grounded answers with provenance, calibrated confidence, and an honest verdict.

HealthWatch has two users: the human who owns the data, and the AI agent acting on their behalf. The same neutral fusion layer serves both — humans get a chat + biometric dashboard; agents get a verifiable `HealthAssertion` (claim, evidence, calibrated confidence, verdict, provenance) over MCP, trust-tiered by who is asking.

Important: HealthWatch is positioned as educational and research, not medical advice. It does not diagnose, prescribe, or auto-act — no ordering, booking, or purchases. Every assertion carries its grounding and an explicit "not medical advice" flag. Raw personal health data (PHI) is never exfiltrated to a third-party agent.

## For humans

- [Home](https://healthwatch.id/): What HealthWatch is — fuse your health data into grounded, cited answers with debate mode and N-of-1 experiments. Invitation only.
- [Pricing](https://healthwatch.id/pricing): Plans and what each tier includes.
- [Sign in](https://healthwatch.id/signin): Invitation code, then Google sign-in.

## For AI agents

- [Agent overview](https://healthwatch.id/for-agents): The human-readable contract — the HealthAssertion envelope, the trust × PHI tiers (T0–T3), and the planned MCP tool surface. Honest status: most tools are spec/preview, not yet live.
- [Machine contract](https://healthwatch.id/.well-known/agent.json): Machine-readable discovery — name, contract version, tier map, planned tools, metering rails, contact.
- [HealthAssertion schema](https://healthwatch.id/.well-known/health-assertion.schema.json): JSON Schema for `HealthAssertion/v1`, the unit of trust every answer is returned as.

## What an answer is

Every HealthWatch answer is a `HealthAssertion/v1` object, not free prose:
a `claim`, the `evidence` behind it (each with a source + trust tier + url), a
`verdict` (one of STRONG_SUPPORT / MODERATE_SUPPORT / WEAK_SUPPORT /
INSUFFICIENT_EVIDENCE / CONTESTED / AGAINST), a calibrated `confidence` in [0,1],
explicit `uncertainties`, walk-it-back `provenance`, and a `not_medical_advice`
flag. The uncertainty that makes a human hesitate is exactly
what an agent needs to act responsibly.

## Trust × PHI tiers (who is asking decides what they see)

- T0 — Public: any agent. Curated KB + PubMed grounding and evidence synthesis. Zero PHI. (MCP, planned.)
- T1 — Aggregate: any paid agent. Cohort statistics with a k-anonymity floor (cells under the threshold are suppressed, not estimated). (MCP, planned.)
- T2 — Compute-to-data: vetted agent + consent. An algorithm runs inside HealthWatch over consented data; only the aggregate result leaves. Raw rows never move. (ACP job + review, planned.)
- T3 — Personal: the user's own delegated agent only. That user's full record, Twin, and N-of-1 — with their revocable consent. (Delegated OAuth, planned.)

## Legal

- [Terms](https://healthwatch.id/terms): Terms of use, including the not-medical-advice and user-responsibility terms.
- [Privacy](https://healthwatch.id/privacy): How personal health data is handled and the PHI boundary guarantees.

## Contact

- Agent access requests and questions: danny.taniwan@gmail.com