Accelerated runs use labeled transport and speech fixtures. Notification fixtures are not proof of watch delivery. Actual emulator receipts are shown separately below. Physical GPS, radio, background operation, sound, vibration and battery performance remain untested.
Evidence levels are scored separately. Emulator notification delivery is measured in the enrolled-watch ledger below.
Baseline and experimental candidate
The candidate adds quality checks and 3-second persistence. It is not enabled in the repeating demo. Brief encounters can be missed. Android accuracy is an estimate, not a guaranteed boundary.
Independent truth vs detector input
Failure record
Interpretation
All qualifying reference encounters remain in the denominator, including outages. Gray gaps are missing or stale observations. Uncertainty is reported separately from confirmed warnings. There is no legal finding.
Repeated samples are correlated. These engineered scenarios do not estimate a population failure rate or physical-watch reliability.
Actual enrolled-watch notification evidence
Queued → app received → rendered → explicit action. Rendering does not prove the wearer noticed. Acknowledgement does not approve assistance. Targets are engineering test values.
Review and responsibility
Privacy failure probes
No imported probe evidence.
Policy propagation
Actual device-applied revision is shown above. Structured human-reviewed rule versions and scoped exceptions are immutable; activation is a separate administrator action. A movement exception never automatically waives proximity.