Nothing is claimed until a traveler has tested it

The gates, in order

Evaluation proceeds only after each preceding gate is passed. This sequencing is the reason the app is not on the App Store, and it is a decision rather than a delay.

1. Software invariants. Boundaries, timing, priority, and health, exercised by deterministic adversarial tests.
2. Measured scenes. Precision, range, latency, and thermal behaviour on a measured obstacle-course benchmark.
3. Engineering walks. Trained staff, ordinary and non-blindfolded, with a spotter.
4. Supervised pilot. Blind and low-vision participants — only after SRC/IRB approval, accessible consent, predefined stopping rules, and orientation-and-mobility supervision.
5. Expanded validation. Unseen routes and an independent team.

Gates one and two are engineering evidence and cannot be reported as participant outcomes. That distinction is enforced throughout the paper.

Why participant work has to be supervised

An assistive travel aid cannot be validated by its authors walking familiar routes. The failure modes that matter appear when a traveler who does not already know what the system is doing encounters a street the system has not seen, and the only way to observe that safely is with an orientation-and-mobility specialist present.

A session pauses if the spotter intervenes, if processing or mount health goes stale, if localization conflicts with the physical corridor, if commands repeatedly contradict each other, if the participant asks to stop, or if a crossing would require the system to grant permission.

What gets measured

The design is within-subjects and counterbalanced. Each participant keeps their usual mobility aid and is measured with it alone against it plus NaviGuider, with a separate training route before any scored one. Primary outcomes:

• Environmental contacts and adjudicated near misses per 100 metres.
• Wrong turns and overshoots.
• Experimenter interventions.
• Successful completion.

Completion time, preferred-speed fraction, false stops, command reversals, and an accessible workload measure are guardrails, not goals — a system that gets you there faster while exhausting you has not helped. Near-miss criteria are fixed before collection and reviewed independently, because the app’s own log supports an observer rather than replacing one.

Field replay: turning walks into experiments

Once a clip recorder existed, judging a tuning change stopped requiring a re-walk. A shadow harness re-runs recorded walks under baseline and candidate configurations and reports every decision divergence. The clip format carries depth, confidence, pose, motion, GPS, route state, guidance, health, events, compressed RGB, and typed perceptual results, joined by frame identity — a record that cannot be joined is counted and refused rather than attached to a neighbouring frame.

It is a powerful paired engineering test with one honest limit: disagreement between two configurations does not tell you which one is correct. Human labels, measured course geometry, or observed outcomes are still necessary, and live evaluation is still necessary because a changed command alters the traveler’s next action and therefore the next sensor reading.

Research Mode

Research Mode is the only path by which camera, depth, or location data leaves the device. It is off by default, it is consent-gated, and it is scoped to a trial. Normal navigation processes sensor data on-device and uploads none of it.

When Research Mode is on, NaviGuider assembles a bundle around a reported or flagged moment and uploads it. A bundle contains:

Because a bundle ties those pieces together by design, they are declared as linked to the research identifier in the app’s privacy manifest. Declaring them unlinked would be false. There is no tracking, no advertising identifier, and no third-party analytics SDK anywhere in the app. The full detail is on the privacy page.

Taking part

We are looking for two groups. Blind and low-vision travelers who already travel independently and are willing to walk supervised routes. And orientation-and-mobility specialists willing to observe, challenge the design, and tell us where the guidance model is wrong.

Email nhellyer6@gmail.com with “research” in the subject. Tell us roughly where you are and how you travel. Participation involves consent to Research Mode for the duration of the trial and nothing beyond it.