# Pulse IQ safety research and validation framework

Research date: 15 October 2026. GuardBand is a working name. This is a research and engineering proposal, not evidence of tested product performance. The accompanying 62-scenario register contains hypotheses and proposed tests, not completed tests. No finite list covers all possible situations.

## Executive finding

A staged wearable safety concept is technically worth investigating, but neither a sequence of three unanswered prompts nor multiple sensor signals establishes a 99% probability of genuine danger. Nonresponses are correlated: a sleeping wearer, disconnected phone, disabled notification channel or inaccessible interface can explain every missed prompt. Repeated prompts may delay a real emergency and produce alarm fatigue. The product should seek fewer unnecessary escalations while tracking missed emergencies, delivery failures and time to help as co-equal outcomes.

Use public wording: **“Research target: up to 99% fewer false escalations against a defined baseline, subject to validation. No performance claim has yet been established.”** Do not use “99% accurate,” “99% real threats,” “prevents assault,” or “guaranteed rescue.” Consider omitting the numeric target from the homepage and explaining it in the research section only.

## Primary-source evidence and engineering implications

1. Apple explicitly documents that fall detection misses some falls and high-impact activity can trigger it. It uses user prompts and an immobility pathway; connectivity is required. This supports explicit error and network disclosures, not copying its timing into a different assault-detection use case. [Apple Fall Detection](https://support.apple.com/en-in/108896).
2. Google's loss-of-pulse research uses multiple physiological/signal gates, then responsiveness prompts, and describes prospective everyday-wear validation. It is a narrowly defined physiological target. Its study does not validate inference of violence from movement or heart rate. It demonstrates the need for separate real-world nuisance-rate and event-detection studies. [Google Research](https://research.google/blog/loss-of-pulse-detection-on-the-google-pixel-watch-3/).
3. Android documents accelerometers as acceleration sensors and gyroscopes as rotational sensors. A compass aids orientation; these are not a standalone reliable latitude/longitude service. Hardware availability varies. Design a supported-device capability matrix rather than assuming arbitrary bands expose their raw sensors. [Android sensors overview](https://developer.android.com/develop/sensors-and-location/sensors/sensors_overview), [motion sensors](https://developer.android.com/develop/sensors-and-location/sensors/sensors_motion).
4. GPS.gov documents degraded positioning near buildings and in obstructed environments. Display fix age and estimated accuracy, and label last-known positions. A precise pin is not proof of precise location. [GPS Accuracy](https://www.gps.gov/gps-accuracy).
5. Apple documents that fit, motion and perfusion can affect optical heart readings. Heart rate or inferred stress is supporting context, not proof of an assault. Missing data should be represented as unknown, not safe. [Apple measurement guidance](https://support.apple.com/en-in/105002).
6. Android requires purposeful, visible background location use and appropriate permissions. An ordinary website cannot provide reliable continuous native background protection. A website demonstration must be identified as a simulation. [Android background location](https://developer.android.com/develop/sensors-and-location/location/background).
7. India's Ministry of Home Affairs describes ERSS-112 as a State-centric system with call-taking and dispatch workflows. That does not establish an open production integration for this project. Police/SHE-team dashboards must be a future partner concept until agreements, routing, verification and operating procedures exist. [MHA ERSS](https://www.mha.gov.in/en/commoncontent/emergency-response-support-system-erss).
8. NIST AI RMF calls for measuring errors and generalizability. A meaningful assurance plan includes deployment monitoring and external validity, not a single laboratory accuracy score. [NIST AI RMF](https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf).

All design recommendations below are our synthesis and proposals; none is an endorsement by these organizations.

## Proposed decision architecture

### Separate three kinds of events

- **Explicit request for help:** a supported button, deliberate configured gesture or another accessible intentional action. Route to the configured response plan without making the user pass a model-confidence test or three verification rounds. Whether a gesture is deliberate requires usability and confusion testing.
- **Possible automatically detected incident:** start a bounded, accessible check-in with a visible deadline and uncertainty-aware incident label. Escalation timing must be derived from a hazard analysis and validated study, not copied arbitrarily from another product.
- **System degradation:** offline phone, removed device, sensor gaps, permissions revoked or battery failure. Show reduced protection and attempt supported recovery. Do not classify these conditions alone as proof of an assault; do not allow them to silently resolve an active incident.

### State machine

Monitoring → candidate detected → check-in → contact notification → acknowledgment/response coordination → resolved.

Independent transitions: explicit SOS can bypass check-in; active incident can become connectivity degraded; failed delivery retries within a bounded policy; resolution sends an update to everyone previously notified. “Delivered,” “read,” “acknowledged,” and “help dispatched” are separate facts. A guardian tapping acknowledge does not establish that the wearer is safe. Arrival at a destination does not close an unresolved emergency. Three repeats must never be treated as three independent confirmations.

A future responder handoff requires a configured legitimate endpoint and jurisdiction-specific operational agreement. Until then, explain the user's own official emergency-call option; do not simulate a real dispatch or imply an authority receives data.

### Signal processing and personalization

Begin with capability discovery, fit/on-body confidence, timestamps, sampling health and battery state. Apply windows of motion features, activity context and sensor-quality indicators; inspect temporal evidence instead of a single spike. A calibrated model may combine available signals but must abstain when unsupported or uncertain. Heart data cannot be required for manual help or used as proof of intent.

Use a stable baseline model from day one for supported features; a month of use is not an evidence-based universal learning threshold. Personalization requires sufficient diverse examples and user-confirmed labels. Never treat every cancelled prompt as a trustworthy negative label: a cancellation can be coerced or accidental. Bound updates, keep a holdout set, version changes, test regressions and support rollback. Workout/travel modes should expire automatically and must preserve explicit SOS. No mode should promise perfect distinction between sport and violence.

## What the 99% target means mathematically

Choose a denominator before collecting data. Let B be baseline false escalations per 1,000 wearer-hours and N be candidate-system false escalations per 1,000 wearer-hours under the same protocol. Relative reduction = 1 − N/B. A 99% reduction means N ≤ 0.01B. It says nothing alone about sensitivity, precision, emergency prevalence or rescue success. Compare matched conditions; disclose the baseline, version, participant population, exposure and uncertainty.

Hypothetical base-rate example, not product results: among 100,000 independent opportunities, assume 100 genuine events, 90% sensitivity and a 1% false-positive rate among non-events. There are 90 true alerts and 999 false alerts, so precision is about 8.3%, despite 99% specificity. Real wearable samples are correlated and “opportunities” must be defined; report event-level and wearer-time metrics as well.

For zero observed false escalations over T wearer-hours, the approximate one-sided 95% Poisson upper rate is 3/T, assuming an appropriate event process. Zero does not prove zero risk. As an illustrative planning calculation, an upper rate below 1 per 10,000 hours would need roughly 30,000 zero-event hours under those assumptions. Participant clustering, rare activities, distribution shift and uncertainty in the baseline require a proper study design beyond this shortcut.

## Evaluation plan and release gates

1. **Define scope:** choose supported form factor, population, jurisdiction, detected event types and claims. Establish a hazard register and ownership. Do not begin with every form factor, every emergency and every demographic.
2. **Bench and integration testing:** clocks, missing data, wear status, permission revocation, battery, duplicate events, cancellation races, reconnection and server outages. Use synthetic recipients; never place test calls to emergency services.
3. **Human factors:** informed-consent studies of gesture discoverability, accidental activation, under-stress interface comprehension, accessibility, safe cancellation, coercion risks and alarm fatigue. Do not stage actual assaults or medically dangerous events.
4. **Safe activity dataset:** consented daily activities across occupations and contexts. Hold out participants, sessions and devices so adjacent windows from one person do not leak across train/test sets. Include motion-confusable negatives and unsupported conditions.
5. **Shadow mode:** no live escalation, transparently labelled; collect real-world exposure, candidate prompts and independently reviewed labels. Account for missing outcomes rather than silently treating them as benign.
6. **Supervised pilot:** only after predeclared safety and usability gates. Opt-in trusted contacts, supported channels, clear runbooks, independent incident review, staged rollback and stopping rules.
7. **Prospective external validation:** freeze model and thresholds before evaluation; test new users and locations. Verify delivery separately from detection. External review is needed before strong public safety claims.
8. **Ongoing monitoring:** model/version drift, subgroup gaps, missed incidents, false prompts, uninstall rates, prompt response times, receiver overload and security incidents.

Predeclare minimum sensitivity and maximum acceptable delay by event category with qualified safety specialists. Do not optimize false positives alone. Report confidence intervals and denominators; if evidence is inadequate for a subgroup, disclose that limitation rather than averaging it away.

### Minimum scorecard

| Metric | Why it matters |
|---|---|
| False candidate prompts / 1,000 wearer-hours | Measures everyday irritation |
| False contact escalations / 1,000 wearer-hours | Measures downstream burden |
| False authority handoffs / exposure, if ever supported | Measures burden on responders |
| Sensitivity and false-negative rate by event category | Shows emergencies missed |
| Precision by real-world event prevalence | Shows what an alert means |
| Median and p95 detection-to-send latency | Finds delay hidden by average speed |
| Delivery and acknowledgment rates, separately | Distinguishes network success from human response |
| Stale-location rate and fix age | Exposes navigation uncertainty |
| Battery/runtime and degraded-state duration | Measures actual availability |
| Cancellation completion and accidental cancellation rate | Measures usability and false reassurance |
| Subgroup metrics by fit, skin tone, movement/accessibility, device | Finds unequal performance |
| Coverage of supported versus unsupported conditions | Prevents silent overgeneralization |

## Privacy, coercion and security requirements

A wearer must understand who can access location and when. Prefer local feature processing; collect only necessary telemetry, with explicit opt-in for research. Use scoped time-limited access, encryption, least-privilege roles, recipient verification, revocation and access logs. Define retention before pilots. A guardian, employer or partner must not silently enable continuous surveillance. Distinguish adult autonomy from any future minor-specific product, which requires separate review.

A secret “duress password” is not automatically safe: if discovered, it can increase danger. Treat it as an unvalidated specialist-reviewed idea, not a promised solution. Password/OTP requirements must never block requesting urgent help. Cancellation authentication must balance account security, accessibility and coercion risk; a correct password is not proof of safety.

No actual location, heart-rate or SOS capture is needed for the public website. Use fictional demonstration data and clearly mark simulated outcomes.

## Product feasibility and credibility

An editable public research website can credibly show the concept, design rationale, limitations, test protocol and development roadmap. It cannot establish that a product exists, a patent was filed/granted, a police partnership exists or a performance target was achieved. Do not display fabricated customers, testimonials, team members, funding, certifications, incorporation identifiers or research results. If patent filing is not documented, say “IP strategy under evaluation,” not “patented” or “patent pending.” Any patentability/public-disclosure decision needs separate qualified advice before confidential enabling details are published.

Propose a phased scope: public concept and simulation; phone-connected proof of concept on explicitly supported hardware; bench validation; consented shadow study; limited opt-in pilot; only then expanded hardware or responder integrations. An original operating system, broad wellness metrics and multiple device forms add independent validation and engineering work; they should remain separate roadmap hypotheses.

## Scenario register

See `scenarios.json`: 62 distinct proposed scenarios covering everyday motion, sensor integrity, location, connectivity, genuine danger, coercion, security and accessibility. Each includes risk, proposed response and proposed test. No entry is a claim that the system already handles that situation.
