Methods — Banlys MFD

Abstract

Banlys MFD is a licensed, versioned package of multi-sensor readings from consumer phones. People turn on research sharing in a Banlys mobile sensing app. When they are sensing, the phone sends short batches: magnetometer (used here as an EMF proxy), motion, gyro, light, barometer when the hardware has one, location rounded to about 100 meters, and — after a short baseline calibration — entropy summaries over rolling windows.

A package is a filtered export, a dictionary, a manifest with checksums, and terms.

Why we collect it

Field sessions are more useful when they stack: same rough protocol, many phones, many places, and flags you can filter on. The hardware is already in the pocket. What we add is consent, a cadence, a baseline, and a versioned file.

These channels are what consumer phones will give us on a repeatable interval. Calibrated sessions on the same device family are the ones that compare cleanly.

How a row gets born

Person installs a Banlys sensing app
  → research sharing stays off until they turn it on
  → they start sensing
  → the phone takes a baseline (~45 seconds, hold still) if they go through that step
  → about every 30 seconds the app may send a batch
  → Banlys stores the batch
  → Saturday we pull a full export
  → a versioned MFD package is built from that export
  → the licensed copy is that package

Paid features and research sharing are separate. Sharing can be turned off in Settings. A batch is sent when sharing is on and sensing is running.

What we collect

Sensors

Values are device-dependent. Same model compares better than mixing every Android together.

Channel	What it is	Notes
`emf` / `magnet`	Magnetometer	Same sensor in the current app
`motion`	Accelerometer-derived	Still vs moving
`gyro`	Gyroscope	Common on current phones
`light`	Ambient light	Scale varies
`pressure`	Barometer	About 22% of readings in this pull
`entropy`	Shannon summaries + a composite	After calibration

Each reading has a device timestamp. Each batch also has a server received time. Keep those two clocks distinct.

Location

Latitude and longitude are rounded to about 100 meters (three decimal degrees). That is enough for regional maps. Some batches have no location; those are not a place.

Calibration and entropy

A ~45 second hold-still baseline lets later readings include entropy. Those batches are marked calibrated.

Entropy is Shannon entropy over a rolling window (60 samples in the live stream, faster than the 30-second research cadence). Values are scaled to 0–1. composite combines the per-channel entropies and is shifted so a quiet baseline sits near the middle of the scale.

If entropy is missing, there was no baseline on that batch.

Quality and instrument fields (schema v2)

Newer clients send these. Older batches may leave them blank.

Field	Role
`research_quality`	`gold` when calibrated, at least 5 readings, and not flagged degenerate; otherwise `standard`
`entropy_degenerate`	Flat or floor windows. Often blank on older rows
`same_location_as_calibration`	Batch still within ~100 m of the baseline
`device_family`	Coarse model token
`charging` / `activity_state`	Useful around the magnetometer and still vs moving
`sample_interval_ms`	Nominal 30000 on current mobile

gold is a starting filter. High-zero sessions can still wear that label, so look at the values.

What a package is

A release is a folder:

data.json — batches for that SKU
dictionary and terms
manifest.json — release_id, counts, filters, checksums
CHECKSUMS.txt

Sample (this pull: 50 batches) is a small slice of the schema. Standard is the research cut of that Saturday’s export. Annual access covers Standard releases published during the license term.

Snapshot — 2026-09-05

Official pull headlines. Platform counts are schema v2.

Batches	5,794
Schema v2	5,266
Calibrated (entropy present)	3,156
Gold (as labeled)	1,703
Readings	36,466
Readings with entropy	21,776
Entropy composite exactly 0	5,493 (25.2%)
Location cells (~0.001°)	749
Calendar days (any / entropy)	78 / 71
v2 Android / iOS	5,217 / 49

About 54% of opt-in batches finished calibration. Most new volume that week was Android.

Suggested cuts

Use calibrated rows when you need entropy.
Prefer same_location_as_calibration = true when the baseline should match the site.
Set aside cells that are almost all zero.
Split by device_family or at least by platform.
Use charging and activity when those fields are present.

entropy_degenerate is blank on most older batches. If the field is missing, work with the numbers you have. A gold / same-place cut is available as a filter on the same corpus.

Provenance

Banlys MFD packages are derived from opt-in research contributions collected through Banlys mobile sensing applications. Packages include multi-sensor readings, coarse location (~100 m), and quality metadata as documented in the data dictionary. They do not include media files or account identifiers.

Sharing is consumer opt-in. Institutions that need a review for secondary use can treat it as consented consumer sensor data and follow their own process. Location is coarse by design; the license still asks that people not be re-identified from it.

The Banlys sensing app on the stores today is JuJu.

Where to get it

Licenses are on the Access page — sample and annual seat. Other licenses are quoted. Live counts sit on Status. If you use these numbers in a write-up, cite the release_id on the manifest and the pull date (2026-09-05).

Descriptor dated 2026-09-11. Numbers from the 2026-09-05 pull.

Collection protocol