49 KiB
Data quality of the Schwarz PAX BLE counter export
Research note · 14 November 2026
Source: data/raw/schwarz-pax.csv (19.4 MB, read-only)
Scope: reliability and privacy limits of 15-minute BLE PAX aggregates from eight Heilbronn locations, 18 June – 7 November 2026 (143 days), plus the legacy density-01 series from October 2025 to March 2026.
Summary
- The export contains 192,464 rows (96,232 density + 96,232 battery), 12 device IDs, 11 names, and 8 usable locations. The 8 current devices all lie on a 0.7 value grid; the legacy
density-01device (863f75b0) is on a 1.0 integer grid. The twodensity-01series are separated by a 35.2-day gap and must not be pooled without an explicit scale break. - The nominal “15-minute” cadence is not a 15:00 wall-clock grid. Most current devices report every 15.27–15.37 min (median), the new
density-01every 16.14 min, and the legacy device every 15.18 min with ~18 s jitter. The new device produces ~89 readings/day, not 96. - Coverage is strongly non-stationary. Median location-month coverage (48 location-months) is 86.1%, but August 2026 falls to 54.2% on average and 23.3% at the worst location. Six of eight locations show a simultaneous multi-week degradation;
density-01andhagenbuchersee-01do not. - All eight locations report in the same 15-minute slot only 28.2% of the time (3,876 / 13,728). The modal count is seven locations (4,117 slots), not eight. Aggregating to 30/60/120 min raises complete eight-location coverage to 63.4% / 69.7% / 71.0%.
- Values are not raw counts. Overall, 5.18% of readings are zero, 10.36% are
<1, and 39.57% are<5. There are no negative or NaN values. The effective positive step is 0.7 for all current devices and 1.0 for the legacy device; the unit and any vendor correction are undocumented. - Raw cross-location Pearson correlation averages 0.58 (range 0.13–0.90), but drops to 0.38 after removing location-specific hour-of-week means and 0.48 at the daily level. Correlation is not evidence of shared unique visitors.
- Battery telemetry is structurally coupled to density telemetry: every density row has a battery value within 60 s, but both disappear together during outages. Levels are coarse and stale (70–89% of consecutive values identical; runs up to 19 days) and do not predict daily counts within a device (−0.41 to 0.08). Across devices, however, August coverage is strongly confounded with battery (r = 0.84; six degraded devices at 4–10% vs two healthy at 41–70%), so battery, device batch, and gateway are collinear and the outage cause cannot be isolated.
- What cannot be measured from these aggregates: MAC-randomization rate, penetration rate, unique-device counts, dwell time, repeat visits, detection radius, or false-positive/false-negative rates. The 0.7 grid and zero fractions are proxies for vendor processing, not direct measurements of randomization or penetration.
1. Data and methods
1.1 File and schema
The CSV is a long-format export with one row per (time, sensor_id, type) observation. Timestamps are ISO-8601 with sub-second precision and UTC offset (+00). The schema is:
| column | dtype | non-null | example |
|---|---|---|---|
time |
datetime64[us, UTC] | 192,464 | 2026-06-12 12:27:15.241799+00:00 |
sensor_id |
string | 192,464 | 00091a5f |
name |
string | 192,464 | hn-goetzentuem-01 |
longitude |
float64 | 192,464 | 9.2148042 |
latitude |
float64 | 192,464 | 49.1399582 |
type |
string | 192,464 | densityData |
value |
float64 | 192,464 | 10.5 |
There are no nulls in any column. type has exactly two values: densityData (96,232 rows) and batteryData (96,232 rows). Density and battery rows are emitted in pairs; battery timestamps are a median 0.033 s after the density timestamp (p99 5.82 s, max 34.3 s).
1.2 Definitions
- Device ID:
sensor_id(12 unique). Location:name, except that the twodensity-01device IDs are treated as one location for the 8-location panel. - Core devices: nine IDs with >1,000 density readings. They comprise eight current locations plus the legacy
density-01device. The eight current device IDs are the “8 usable sensors”; the legacy ID is retained for the scale-break analysis but excluded from the panel. - Episodic devices: three IDs with 5–73 readings (
hn-bollwerksturm-01,maker-fair,niklas-01). They are test/pilot deployments and are excluded from the panel. - Panel window:
2026-06-18 00:00 UTCto2026-11-08 00:00 UTC(exclusive), i.e. 143 days / 13,728 15-minute slots. This starts at the first full day on which all eight current locations are live, avoiding the partial deployment days of 12–17 June. - Slot: UTC timestamp floored to 15 minutes (
HH:00,HH:15,HH:30,HH:45). Multiple readings in one slot are averaged (rare). - Local time:
Europe/Berlinfor hourly/weekday profiles. The DST transition on 25 October 2026 is handled by the timezone library; one local day has 25 hours. - Coverage: share of the 13,728 panel slots with at least one reading, not share of expected readings. Because the native cadence is 15.27–16.14 min, 96 slots/day is an upper bound that most devices cannot reach even when healthy. A second metric, coverage vs cadence ceiling, divides observed slot coverage by the maximum attainable coverage given the median cadence (
100 × 15 / median_interval_min).
1.3 Software and reproducibility
All computations were done with Python 3.12, pandas 3.1.1 and matplotlib 3.11.2. The analysis script is paper/research/scripts/01_data_quality.py; it reads the raw CSV read-only and writes tables to paper/research/outputs/ and figures to paper/research/figures/. Run:
uv run --with pandas --with matplotlib python paper/research/scripts/01_data_quality.py
The full per-device tables are in the CSV outputs cited below; the main text shows the values needed for the paper.
2. Inventory and time range
Overall range: 2025-10-16 21:17:59 UTC → 2026-11-07 23:51:57 UTC (387.1 days).
Panel range: 2026-06-18 → 2026-11-07 (143 days).
| device_id | name | role | density readings | first density timestamp | last density timestamp |
|---|---|---|---|---|---|
863f75b0 |
density-01 (legacy) | core | 8,075 | 2025-10-16 21:17 | 2026-03-12 18:28 |
f8c19730 |
density-01 (new) | core | 17,585 | 2026-04-16 23:25 | 2026-11-07 23:48 |
00091a5f |
hn-goetzentuem-01 | core | 9,952 | 2026-06-12 12:27 | 2026-11-07 23:39 |
e52c55f3 |
hn-hagenbuchersee-01 | core | 12,345 | 2026-06-12 12:02 | 2026-11-07 23:46 |
0f8519cf |
hn-hagenbuchersee-02 | core | 10,099 | 2026-06-12 12:32 | 2026-11-07 23:51 |
8d7eef22 |
hn-marahaus-01 | core | 9,254 | 2026-06-12 12:36 | 2026-11-07 23:49 |
d4cce74b |
hn-neckarufer-01 | core | 9,527 | 2026-06-12 12:21 | 2026-11-07 23:49 |
ba71377f |
hn-soleo-01 | core | 10,327 | 2026-06-12 11:52 | 2026-11-07 23:32 |
f4928534 |
hn-stadtgalerie-01 | core | 8,982 | 2026-06-12 12:11 | 2026-11-07 23:26 |
d3304b39 |
hn-bollwerksturm-01 | episodic | 5 | 2026-06-12 11:19 | 2026-08-13 16:01 |
2aadbeed |
maker-fair | episodic | 73 | 2026-08-13 16:07 | 2026-08-16 15:06 |
062fe216 |
niklas-01 | episodic | 8 | 2026-05-01 17:19 | 2026-05-01 19:06 |
Important deployment artefacts.
- The first hn-* reading for seven locations is a single test message on 12 June 2026, followed by a 4–5 day gap until 16–17 June. The panel therefore starts on 18 June.
- The
maker-fairdevice is at 9.77°E, 52.38°N (Hannover), roughly 350 km from the Heilbronn cluster (9.21–9.22°E, 49.14–49.15°N). Its 73 readings over three days (mean 117.8) are a separate deployment, not a Heilbronn location. hn-bollwerksturm-01has five readings over 62 days, including a 58-day gap; it is effectively non-operational.- The two
density-01IDs share the same name and coordinates within ~4 m. They are a hardware replacement, not two simultaneous sensors.
3. Value quantization and the density-01 scale discontinuity
This is the most important unit-level finding. The values are not raw counts, and the counting grid changed at the hardware replacement.
3.1 Per-device value grid
“Multiple of 0.7” means the value divided by 0.7 is an integer within floating-point tolerance. “Integer” means the value equals its rounded value. The two sets overlap at 0, 7.0, 14.0, 21.0, … because every tenth 0.7 step is an integer. In this export sensor_id is the device ID, so this per-device table is also the per-sensor table; the only location with two device IDs is density-01.
| device_id | period | n | min–max | unique values | min positive | effective grid | % on 0.7 grid | % integer | % both | % integer only | % neither |
|---|---|---|---|---|---|---|---|---|---|---|---|
f8c19730 |
current | 17,585 | 0–63.0 | 84 | 0.7 | 0.7 | 100.00 | 12.72 | 12.72 | 0.00 | 0.00 |
00091a5f |
current | 9,952 | 0–210.7 | 152 | 0.7 | 0.7 | 100.00 | 9.90 | 9.90 | 0.00 | 0.00 |
e52c55f3 |
current | 12,345 | 0–292.6 | 206 | 0.7 | 0.7 | 100.00 | 11.71 | 11.71 | 0.00 | 0.00 |
0f8519cf |
current | 10,099 | 0–200.9 | 202 | 0.7 | 0.7 | 100.00 | 8.30 | 8.30 | 0.00 | 0.00 |
8d7eef22 |
current | 9,254 | 0–194.6 | 140 | 0.7 | 0.7 | 100.00 | 7.64 | 7.64 | 0.00 | 0.00 |
d4cce74b |
current | 9,527 | 0–255.5 | 192 | 0.7 | 0.7 | 100.00 | 8.72 | 8.72 | 0.00 | 0.00 |
ba71377f |
current | 10,327 | 0–111.3 | 107 | 0.7 | 0.7 | 100.00 | 12.34 | 12.34 | 0.00 | 0.00 |
f4928534 |
current | 8,982 | 0.7–137.2 | 168 | 0.7 | 0.7 | 100.00 | 6.32 | 6.32 | 0.00 | 0.00 |
d3304b39 |
episodic | 5 | 6.3–16.1 | 5 | 6.3 | 0.7 | 100.00 | 20.00 | 20.00 | 0.00 | 0.00 |
2aadbeed |
episodic | 73 | 9.1–165.2 | 57 | 9.1 | 0.7 | 100.00 | 13.70 | 13.70 | 0.00 | 0.00 |
062fe216 |
episodic | 8 | 2.1–6.3 | 3 | 2.1 | 0.7 | 100.00 | 0.00 | 0.00 | 0.00 | 0.00 |
863f75b0 |
legacy | 8,075 | 0–28.0 | 27 | 1.0 | 1.0 | 35.36 | 100.00 | 35.36 | 64.64 | 0.00 |
Findings.
- All 11 non-legacy devices (8 current + 3 episodic) report 100.00% multiples of 0.7. For the eight current panel devices, the minimum positive value is 0.7 and the smallest observed spacing is 0.7; the episodic devices lie on the same grid but their observed minima are higher (2.1–9.1) because they have very few readings.
- The legacy
density-01reports 100% integers and 0% non-integer values; its minimum positive value is 1.0 and its smallest spacing is 1.0. Its 35.36% “multiple of 0.7” share is coincidental: those are the integers 0, 7, 14, 21, 28, which happen to be divisible by 7. - No device uses 0.1 or 0.5 as its finest grid. All values are representable as multiples of 0.1 because 0.7 = 7 × 0.1, and 15.7–23.3% of current readings are also multiples of 0.5 because every fifth 0.7 step (3.5, 7.0, 10.5, …) is a 0.5 multiple. Those are subsets of the 0.7 grid, not separate grids. The overall “multiple of 0.1” share is 100% to floating-point tolerance, which is an artefact of decimal representation, not evidence of 0.1 resolution.
- “Occasional exact integers” are not a separate mode. Current devices hit an integer only when the 0.7-grid index is a multiple of 10 (0, 7.0, 14.0, …). The observed 6.3–12.7% integer share is consistent with that overlap.
- Overall, 94.58% of all density readings are multiples of 0.7 and 17.63% are integers. 0.00% lie on neither grid.
- The 0.7 factor is undocumented. It is consistent with a vendor correction/scaling factor (possibly related to MAC-randomization adjustment) or a unit change, but the export contains no metadata to distinguish these. Do not describe values as “devices” or “raw counts.”
Consistency check: all eight usable (current) sensors are 100% on the 0.7 grid; the legacy density-01 is the only 1.0-grid device; no device uses 0.5 or 0.1 as its base grid. The 0.5 share is a subset of the 0.7 grid, and the 0.1 share is a decimal-representation artefact.
3.2 density-01: two device IDs, two scales, 35.2-day gap
| series | n | first | last | min | max | mean | median | p95 | p99 | zero % | unique | grid |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
863f75b0 (legacy) |
8,075 | 2025-10-16 21:17:59 | 2026-03-12 18:28:06 | 0 | 28.0 | 3.008 | 2.0 | 11.0 | 16.26 | 29.77 | 27 | 1.0 |
f8c19730 (new) |
17,585 | 2026-04-16 23:25:31 | 2026-11-07 23:48:49 | 0 | 63.0 | 9.741 | 7.7 | 28.0 | 38.50 | 5.87 | 84 | 0.7 |
- The gap between the last legacy reading and the first new reading is 35.2065 days (3,041,845.5 s). There is no overlap, so a conversion factor cannot be estimated empirically.
- The legacy range is 0–28 with 27 unique values; the new range is 0–63.0 with 84 unique values. The new maximum 63.0 is exactly 90 × 0.7; the legacy maximum 28 is 28 × 1.0. This is a scale/quantization discontinuity, not just a change of device ID.
- The level shift is large: mean 3.008 → 9.741 (×3.24) and median 2.0 → 7.7 (×3.85). This is not explained by the 0.7 factor alone: if the new values were 0.7 × raw counts, the numeric values would be lower, not higher, for the same raw count. The periods also differ by season (winter vs summer) and possibly by installation/sensitivity (legacy zero share 29.8% vs new 5.9%). The safe conclusion is: the two series are not on the same scale and cannot be pooled or compared directly; the cause of the level difference is confounded.
- The legacy device also has far worse coverage (Section 5): 56.5% of its active 15-minute slots, 417 gaps >1 h, and 29.8% zeros. Treat it as a separate, lower-quality series.
Figure 1. Left: per-device share of readings on the 0.7 grid, integer grid, both, or neither. Right: reported value ladders for the two density-01 devices (0–7 zoom), showing the 1.0 vs 0.7 spacing.
Figure 2. Daily mean density for the legacy (863f75b0, 1.0 grid) and new (f8c19730, 0.7 grid) density-01 devices. The shaded band is the 35.2-day gap with no data; the two series do not overlap, so no empirical conversion factor can be estimated.
3.3 Integrity checks
| check | count |
|---|---|
| exact duplicate rows | 0 |
duplicate (time, sensor_id, type) |
0 |
duplicate (time, sensor_id) |
0 |
| density values negative | 0 |
| density values NaN | 0 |
| battery values negative | 0 |
| battery values >100 | 0 |
| battery values NaN | 0 |
| core inter-reading intervals <60 s | 113 |
- There are no duplicate timestamps on the natural key
(time, sensor_id, type)and no duplicate(time, sensor_id)pairs. The 113 sub-minute intervals are near-duplicates, not exact duplicates: 106 belong to the legacydensity-01device, 6 tostadtgalerie, 1 tomarahaus. The median spacing is 7.0 s, most carry identical values, and the maximum is 53.3 s. These look like repeated publications/retries; deduplicating by keeping one reading per 60 s is advisable. - No negative or NaN values in either series. Battery values are bounded 0–100; density has no declared upper bound and reaches 292.6.
- There is one duplicate sensor name (
density-01) but no duplicatesensor_ids. The name collision is the hardware replacement, not a duplicated device.
4. Cadence and time alignment
The 15-minute grid is nominal, not exact. Each device has its own timer and phase, and the timestamps carry second-level jitter.
| device | kind | median interval (s) | mean (s) | SD (s) | IQR (s) | p05–p95 (s) | within 900 ± 5 s | within 900 ± 30 s | 14–17 min share |
|---|---|---|---|---|---|---|---|---|---|
863f75b0 |
legacy | 910.86 | 916.35 | 18.03 | 909.85–917.25 | 903.10–958.19 | 12.6% | 89.1% | 72.6% |
f8c19730 |
new | 968.52 | 968.95 | 2.41 | 968.02–969.02 | 967.19–974.44 | 0.0% | 0.01% | 96.6% |
e52c55f3 |
current reference | 916.08 | 916.77 | 3.44 | 915.61–916.80 | 914.67–922.11 | 0.4% | 99.3% | 97.4% |
Across the core devices, 72.6–97.4% of intervals fall in the 14–17 min band and 2.0–12.9% in the 28–34 min band; the latter are missed single ticks, not a slower timer. Longer intervals are dropouts.
- Legacy
863f75b0: the near-cadence median is 910.86 s = 15.181 min, with SD 18.0 s and IQR 909.85–917.25 s. Only 12.6% of intervals are within 5 s of 900 s; 89.1% are within 30 s. The first intervals are 1021.19 s, 942.45 s, 911.49 s, then a 2089.20 s gap (missed ticks), then 912.77 s. This is a ~15.2-minute timer with tens-of-seconds jitter and frequent missed transmissions, not a clean 15:00 clock. - New
density-01f8c19730: median 968.52 s = 16.142 min, SD 2.41 s, IQR 968.02–969.02 s. It is extremely regular but 7.6% slower than 15 minutes; only 0.01% of intervals are within 30 s of 900 s. At 16.14 min it produces ~89 readings/day, so 96 slots/day is unreachable. Its observed panel coverage (88.4%) is 95.2% of its cadence ceiling (92.9%). - Current reference
e52c55f3: median 916.08 s = 15.268 min, SD 3.44 s, 99.3% within 900 ± 30 s. The other current devices are similar (15.27–15.37 min). - Consequence: flooring to
HH:00/15/30/45is an approximation. A device’s reading can sit anywhere inside the bin, and two devices matched to the same slot can be up to ~15 minutes apart. This attenuates cross-sensor correlation and makes the 15-minute complete-case panel sparse. Section 5 quantifies the remedy: 30–60-minute bins.
Figure 3. Left: distribution of inter-reading intervals for core devices (≤60 min), with peaks at ~15 and ~30 min and a small <1 min retry/duplicate bar. Right: median cadence with IQR; annotations give the share of ~30-minute intervals (missed single ticks).
5. Coverage and missingness
5.1 Per-device coverage and gaps
Panel = 18 Jun–7 Nov 2026 (143 days, 13,728 slots). “Gaps” are inter-reading intervals inside the panel. The legacy density-01 has no panel rows (it ended in March); its active-span figures are shown for comparison.
| device | first | last | n | active-span coverage | panel coverage | readings/day (panel) | median cadence (min) | cadence ceiling | coverage vs ceiling | time in gaps >1 h | gaps >1 h | gaps >1 d | max panel gap (h) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| goetzentuem-01 | 2026-06-12 | 2026-11-07 | 9,952 | 69.8% | 72.3% | 69.4 | 15.36 | 97.7% | 74.0% | 18.6% | 177 | 4 | 42.6 |
| hagenbuchersee-02 | 2026-06-12 | 2026-11-07 | 10,099 | 70.8% | 72.7% | 69.8 | 15.30 | 98.1% | 74.2% | 19.9% | 119 | 5 | 82.3 |
| marahaus-01 | 2026-06-12 | 2026-11-07 | 9,254 | 64.9% | 67.2% | 64.5 | 15.28 | 98.2% | 68.4% | 25.4% | 73 | 9 | 152.3 |
| soleo-01 | 2026-06-12 | 2026-11-07 | 10,327 | 72.4% | 74.4% | 71.4 | 15.28 | 98.2% | 75.8% | 17.8% | 116 | 4 | 58.1 |
| neckarufer-01 | 2026-06-12 | 2026-11-07 | 9,527 | 66.8% | 69.2% | 66.4 | 15.31 | 98.0% | 70.6% | 22.1% | 176 | 4 | 42.0 |
| hagenbuchersee-01 | 2026-06-12 | 2026-11-07 | 12,345 | 86.6% | 89.1% | 85.5 | 15.27 | 98.2% | 90.7% | 7.2% | 36 | 6 | 25.0 |
| stadtgalerie-01 | 2026-06-12 | 2026-11-07 | 8,982 | 63.0% | 65.2% | 62.6 | 15.37 | 97.6% | 66.8% | 23.4% | 149 | 4 | 141.8 |
| density-01 (new) | 2026-04-16 | 2026-11-07 | 17,585 | 89.3% | 88.4% | 84.9 | 16.14 | 92.9% | 95.2% | 0.5% | 13 | 0 | 4.6 |
| density-01 (legacy) | 2025-10-16 | 2026-03-12 | 8,075 | 56.5% | — | — | 15.18 | 98.7% | — | — | 417 | 3 | 58.7 |
Reading the table. For the legacy density-01 row, gap counts and maximum gap are over its full active span (it has no panel rows); all other rows are panel-based. The distribution metrics requested for each sensor (zero fraction, <1%, <5%, mean, median, p95, p99, max) are in the table in Section 6.
- “Active-span coverage” uses each device’s own first–last span; “panel coverage” uses the common 143-day window.
- “Cadence ceiling” is the maximum slot coverage attainable at the device’s median cadence (
100 × 15 / median_min). A device at 16.14 min cannot exceed 92.9% of 96 daily slots even with no outages. - The legacy device’s 417 gaps >1 h and 29.8% zeros make it unsuitable as a continuous series. It is also the only 1.0-grid device.
- For the eight current devices, roughly 18–25% of panel time falls inside gaps >1 h for six locations (17.8% at
soleo);hagenbuchersee-01is 7.2% anddensity-010.5%. Gaps >1 h are not just single missed 15-minute ticks: they number 36–177 per device and include many 1–6 h intervals (seeoutputs/large_gaps.csv). - The maximum panel gap is 152.3 h (6.3 days) at
marahaus-01, followed by 141.8 h atstadtgalerie-01and 82.3 h athagenbuchersee-02.density-01(new) never exceeds 4.6 h.
5.2 Monthly coverage and the August outage
| month | density-01 | goetzentuem | hagenbuchersee-01 | hagenbuchersee-02 | marahaus | neckarufer | soleo | stadtgalerie | mean |
|---|---|---|---|---|---|---|---|---|---|
| 2026-06 (partial) | 86.7 | 79.7 | 90.1 | 94.5 | 78.0 | 77.0 | 92.1 | 73.6 | 83.9 |
| 2026-07 | 87.1 | 86.5 | 95.2 | 88.5 | 85.4 | 83.2 | 90.7 | 78.9 | 87.0 |
| 2026-08 | 84.8 | 52.0 | 94.1 | 49.1 | 28.9 | 46.9 | 54.6 | 23.3 | 54.2 |
| 2026-09 | 90.6 | 57.2 | 66.8 | 51.0 | 56.0 | 52.7 | 53.0 | 66.9 | 61.8 |
| 2026-10 | 91.0 | 86.8 | 97.0 | 89.2 | 88.6 | 86.2 | 88.0 | 84.1 | 88.8 |
| 2026-11 (partial) | 92.7 | 86.8 | 98.2 | 87.5 | 89.3 | 86.0 | 88.7 | 82.6 | 89.0 |
This is not a stationary missingness process. July and October are healthy (mean 87–89%), while August collapses to 54.2% and September to 61.8%. Six of eight locations degrade together in August, but density-01 (84.8%) and hagenbuchersee-01 (94.1%) remain healthy. This split is strongly confounded with battery: the six degraded devices had August median battery levels of 4–10%, while the two healthy devices had 41% and 70%; across the eight devices (n = 8), August coverage and August median battery correlate at r = 0.84. That is consistent with a power/battery-related failure, but it is not conclusive: four of the six still report in October at 1–3% while two show battery resets to ~32–36% (likely replacement/charging), the battery gauge is coarse/stale, and battery is collinear with device batch, gateway, and installation. The export has no status/error codes, so the cause cannot be isolated.
5.3 All-eight complete cases are the exception
The panel has 13,728 15-minute slots. The distribution of the number of locations reporting per slot is:
| locations reporting | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 |
|---|---|---|---|---|---|---|---|---|---|
| slots | 69 | 594 | 1,133 | 721 | 615 | 769 | 1,834 | 4,117 | 3,876 |
| share | 0.5% | 4.3% | 8.3% | 5.3% | 4.5% | 5.6% | 13.4% | 30.0% | 28.2% |
- All eight locations report in only 3,876 / 13,728 slots (28.2%), or 28.4% of the 13,659 slots with any data. The modal count is seven, not eight. Even in July (a healthy month), all-eight is only 36.5%; in August it is 6.7%; in October it is 40.3%.
- Under independence, the product of the eight per-location coverage rates would predict only 1,281.5 all-eight slots (9.3%). The observed 3,876 is 3.0× higher. Presence is therefore positively correlated across devices: they tend to be up together and down together, consistent with shared platform outages.
- Aggregation fixes much of this. Re-binning to 30 minutes raises all-eight coverage to 4,350 / 6,864 = 63.4%; 60 minutes to 2,393 / 3,432 = 69.7%; 120 minutes to 1,218 / 1,716 = 71.0%. Pairwise mean correlation changes little (0.58 → 0.60 → 0.62 → 0.63), so 30–60-minute bins are a reasonable default for complete-case cross-sensor analysis.
| bin | expected bins | bins with any data | all-eight bins | all-eight % | mean pairwise r | median CV |
|---|---|---|---|---|---|---|
| 15 min | 13,728 | 13,659 | 3,876 | 28.2% | 0.58 | 0.66 |
| 30 min | 6,864 | 6,854 | 4,350 | 63.4% | 0.60 | 0.64 |
| 60 min | 3,432 | 3,429 | 2,393 | 69.7% | 0.62 | 0.62 |
| 120 min | 1,716 | 1,715 | 1,218 | 71.0% | 0.63 | 0.60 |
5.4 Missingness is not MCAR
For each location, we compared the median daily mean density on high-coverage days (≥80% of 96 slots) with low-coverage days (<50%). The ratio (low/high) ranges from 0.38 to 1.49 (median 0.69): for six of eight locations, low-coverage days also have lower density (38–77% of high-coverage days). This is confounded with the August holiday period, which has both lower footfall and the outages. Either way, missingness is not missing completely at random with respect to the measured quantity. Complete-case seasonal means will be biased, and the direction differs by location. Missingness by local hour is relatively flat (roughly 25–40% for most devices), so there is no simple night-time shutdown pattern.
Figure 4. Daily share of the 96 possible 15-minute slots with a reading. The dark August–September band is a systemic multi-location outage; note that density-01 and hagenbuchersee-01 remain mostly green.
Figure 5. Daily mean density per location. Gaps are days with <80% slot coverage (partial days excluded, not interpolated). The series show a genuine multi-day high-traffic period on 18–20 June, the August outage, and a broad seasonal decline into November.
6. Value distribution, zero inflation, and low counts
| device | zero % | <1 % |
<5 % |
mean | median | p95 | p99 | max |
|---|---|---|---|---|---|---|---|---|
| goetzentuem-01 | 2.97 | 11.28 | 37.82 | 11.12 | 8.40 | 29.40 | 53.90 | 210.70 |
| hagenbuchersee-02 | 0.69 | 3.03 | 30.42 | 19.28 | 11.20 | 64.40 | 96.61 | 200.90 |
| marahaus-01 | 0.01 | 1.76 | 33.93 | 13.73 | 9.10 | 38.50 | 58.10 | 194.60 |
| soleo-01 | 4.77 | 9.66 | 32.13 | 10.58 | 8.40 | 28.00 | 39.20 | 111.30 |
| neckarufer-01 | 0.07 | 0.50 | 23.24 | 19.20 | 11.90 | 61.60 | 86.80 | 255.50 |
| hagenbuchersee-01 | 5.56 | 14.34 | 45.69 | 12.97 | 6.30 | 49.70 | 85.40 | 292.60 |
| stadtgalerie-01 | 0.00 | 9.26 | 40.79 | 21.14 | 8.40 | 67.20 | 88.90 | 137.20 |
| density-01 (new) | 5.87 | 13.20 | 39.68 | 9.74 | 7.70 | 28.00 | 38.50 | 63.00 |
| density-01 (legacy) | 29.77 | 29.77 | 77.78 | 3.01 | 2.00 | 11.00 | 16.26 | 28.00 |
- Overall: 5.18% zeros, 10.36%
<1, 39.57%<5. The distribution is heavy-tailed and zero-inflated; a log1p or negative-binomial treatment is more appropriate than Gaussian. - A
<1reading is not necessarily zero: on the 0.7 grid it means 0 or 0.7.stadtgalerie-01has no zeros (minimum 0.7) yet 9.26%<1, because all of those are 0.7 readings. - The legacy
density-01is an outlier: 29.8% zeros and 77.8%<5, consistent with a lower-sensitivity or differently scaled deployment. It should not be pooled with the new series. - There are no negative or NaN values. Density has no declared upper bound; the observed maximum is 292.6 at
hagenbuchersee-01on 2026-08-14 19:19. The distribution is not truncated at 100 or any other ceiling.
Figure 6. Left: histogram of density values (0.7 bin width; log y). Middle: CCDF on log-log axes. Right: per-device boxplots on a symlog scale. The legacy device is visibly lower and more zero-inflated; current devices share a similar heavy-tailed shape.
7. Cross-location correlation and ratio stability
7.1 Pairwise correlation at matching 15-minute slots
The matrix uses the 8-location panel, pairwise complete observations. Each cell is Pearson r; N ranges from 7,261 (neckarufer–stadtgalerie) to 10,827 (density-01–hagenbuchersee-01). The diagonal is 1.00.
| location | density-01 | goetzentuem | hagenbuchersee-01 | hagenbuchersee-02 | marahaus | neckarufer | soleo | stadtgalerie |
|---|---|---|---|---|---|---|---|---|
| density-01 | 1.00 | 0.60 | 0.64 | 0.69 | 0.74 | 0.73 | 0.49 | 0.44 |
| goetzentuem | 0.60 | 1.00 | 0.75 | 0.63 | 0.78 | 0.72 | 0.55 | 0.37 |
| hagenbuchersee-01 | 0.64 | 0.75 | 1.00 | 0.83 | 0.79 | 0.87 | 0.48 | 0.15 |
| hagenbuchersee-02 | 0.69 | 0.63 | 0.83 | 1.00 | 0.73 | 0.90 | 0.45 | 0.13 |
| marahaus | 0.74 | 0.78 | 0.79 | 0.73 | 1.00 | 0.80 | 0.53 | 0.41 |
| neckarufer | 0.73 | 0.72 | 0.87 | 0.90 | 0.80 | 1.00 | 0.52 | 0.19 |
| soleo | 0.49 | 0.55 | 0.48 | 0.45 | 0.53 | 0.52 | 1.00 | 0.40 |
| stadtgalerie | 0.44 | 0.37 | 0.15 | 0.13 | 0.41 | 0.19 | 0.40 | 1.00 |
Robustness to transformation and time scale:
| estimator | mean off-diagonal r | min | max | N |
|---|---|---|---|---|
| Pearson, 15-min raw | 0.582 | 0.130 | 0.904 | 7,261–10,827 pairs |
| Spearman, 15-min | 0.678 | 0.438 | 0.871 | same pairs |
| Pearson, log1p | 0.670 | 0.435 | 0.862 | same pairs |
| Pearson, hour-of-week mean removed | 0.385 | 0.031 | 0.785 | same pairs |
| Pearson, daily means (69 high-coverage days) | 0.477 | −0.063 | 0.937 | 69 days |
- Raw Pearson is 0.58 on average but ranges from 0.13 (
hagenbuchersee-02vsstadtgalerie) to 0.90 (hagenbuchersee-02vsneckarufer). Spearman and log1p are higher (0.67–0.68), indicating that raw Pearson is attenuated by spikes and zero inflation; rank/log transforms capture the shared ordering better. - Removing location-specific hour-of-week means drops the average to 0.385. At the daily level, the average is 0.477 (69 days with all eight locations at ≥80% coverage). So roughly 15–23% of variance is shared after removing the common weekly profile, versus ~34% raw (0.58²). The rest is location-specific level, semantics, and noise.
- The low pair is interpretable:
stadtgalerie(the name and profile suggest a shopping center) has a midday/afternoon peak and Sunday near-closure;hagenbuchersee-02(a lake location) has an evening peak. Their raw r = 0.13 is not a sensor failure; it is a different population/flow. - Correlation is not deduplication. Two sensors can correlate because they see the same city rhythm, weather, or events, not because they see the same devices. With aggregates only, unique-visitor overlap is not identifiable.
Figure 7. Left: raw Pearson. Middle: Spearman. Right: Pearson after removing each location’s hour-of-week mean. The third panel is the conservative view: shared sub-weekly structure is much weaker than the raw matrix suggests.
7.2 Cross-sensor dispersion and ratio stability
- Raw coefficient of variation across the eight locations at the same 15-minute slot: median 0.66, p90 1.05, p95 1.18. This is dominated by the different mean levels of the locations.
- Scale-free share CV (each location’s share of the eight-location total at that slot): median 0.216, p95 0.463. Even after removing level differences, the cross-sectional composition varies substantially from slot to slot.
- Daily share stability: over the 69 days on which all eight locations have ≥80% coverage, the daily share CV per location ranges from 0.159 (
neckarufer) to 0.508 (soleo); the median is 0.216. - Pairwise ratio stability: the median SD of the daily log ratio between two locations is 0.399 (a 1-SD multiplicative change of e^0.399 = 1.49×); the p90 is 0.711 (2.04×). A fixed conversion factor between two sensors is not supported. Ratios drift with season, events, and location-specific behavior.
7.3 Hourly and weekday profiles
| location | peak hour (local) | weekend/weekday ratio |
|---|---|---|
| density-01 | 19 | 1.18 |
| goetzentuem-01 | 19 | 1.12 |
| hagenbuchersee-01 | 20 | 1.24 |
| hagenbuchersee-02 | 21 | 1.19 |
| marahaus-01 | 19 | 1.17 |
| neckarufer-01 | 20 | 1.14 |
| soleo-01 | 19 | 0.98 |
| stadtgalerie-01 | 16 | 0.81 |
- Seven locations peak in the late afternoon/evening (16:00–21:00);
stadtgaleriepeaks at 16:00 and drops sharply after 20:00, consistent with a shopping-center closing time. - Weekend/weekday ratios are 1.12–1.24 for the outdoor/leisure locations, 0.98 for
soleo, and 0.81 forstadtgalerie. The mall’s Sunday value is only ~0.3 of its mean, i.e. near-closure, while the outdoor locations peak on Saturday (~1.4 of mean). This heterogeneity is a semantic difference, not a sensor error. - Mean pairwise correlation of the normalized hourly shapes is 0.79, so the broad day/evening rhythm is shared even where the levels and weekend behavior differ.
Figure 8. Normalized hourly (left) and weekday (right) profiles. The mall (stadtgalerie, grey) is the clear outlier: midday peak and Sunday near-closure.
8. Battery telemetry
Battery rows are emitted with density rows, not independently. Every one of the 96,232 density rows has a battery value within 60 s (100%; median offset 0.033 s, p99 5.82 s). But during a device outage both series disappear, so battery missingness is structurally identical to density missingness. In the panel, the eight active devices contribute 82,170 paired rows out of 109,824 expected device-slots (74.8%); the other 25.2% have neither density nor battery. “0% battery missing conditional on a density row” is therefore not evidence of reliable battery telemetry.
| device | first → last (%) | median (%) | share <20% |
consecutive-equal share | longest stale run (days) | corr(battery, daily readings) |
|---|---|---|---|---|---|---|
| goetzentuem-01 | 99 → 0 | 4 | 100% | 89% | 11.5 | 0.02 |
| hagenbuchersee-02 | 92 → 27 | 8 | 53% | 85% | 19.1 | −0.16 |
| marahaus-01 | 52 → 87 | 7 | 85% | 81% | 10.8 | −0.06 |
| soleo-01 | 94 → 26 | 8 | 51% | 83% | 10.0 | −0.04 |
| neckarufer-01 | 100 → 11 | 8 | 89% | 84% | 11.2 | 0.08 |
| hagenbuchersee-01 | 92 → 43 | 57 | 16% | 74% | 3.6 | −0.02 |
| stadtgalerie-01 | 99 → 75 | 4 | 93% | 88% | 11.6 | 0.07 |
| density-01 (new) | 89 → 11 | 46 | 9% | 70% | 3.7 | −0.41 |
| density-01 (legacy) | 44.9 → 69.4 | 31.9 | 7% | 52% | 0.5 | — |
- Overall battery median is 17%, with 51.3% of readings below 20% and 1.6% at 0%. Values are bounded 0–100.
- The current devices are coarse and stale: 70–89% of consecutive battery readings are identical, and the longest constant run is 3.6–19.1 days. Values can also increase (e.g. 4 → 36, 6 → 74), indicating battery replacement, charging, or gauge reset rather than monotonic discharge.
- Within a device, battery does not predict availability. Daily correlation between battery level and daily reading count ranges from −0.41 to 0.08; several devices keep reporting for months at 1–4%.
- At the reading level, corr(battery, density) ranges from −0.14 to 0.42 across the eight current devices (0.53 for the Hannover
maker-fairpilot), again with no consistent sign. - Across devices, August coverage is strongly confounded with battery. The six degraded locations had August median battery 4–10%, the two healthy locations 41% and 70%; the cross-device correlation is r = 0.84 (n = 8; Figure 9, right). But four of the six still report in October at 1–3% while two show battery resets to ~32–36%, so a simple threshold does not explain the outage. Battery, device batch, gateway, and installation are collinear; do not read the association as causal.
- The legacy
density-01battery values are fractional and increase over the deployment (44.9 → 69.4), another firmware difference. Its battery series is not comparable to the integer-percent current devices. - Interpretation: battery telemetry here is a coarse gauge, not a reliable power-state or duty-cycle measurement. It cannot be used to explain missingness cleanly, and it cannot be used to infer scanning duty cycle or power-saving behavior.
Figure 9. Left: daily median battery per device. Middle: distribution of paired battery readings. Right: August slot coverage vs August median battery; the six low-battery devices are also the six degraded ones (r = 0.84). The comb pattern in the middle panel is integer-percent quantization; the long low plateaus are stale values, not necessarily a linear discharge.
9. Duplicates, spikes, flatlines, and dropouts
9.1 Duplicates
- Exact duplicate rows: 0. Duplicate
(time, sensor_id, type): 0. Duplicate(time, sensor_id): 0. The export is unique on its natural key. - Near-duplicates: 113 inter-reading intervals <60 s, of which 106 are the legacy
density-01. The median spacing is 7.0 s (range 5.1–53.3 s), and most pairs carry identical values. This looks like repeated publication/retry, especially in the legacy firmware. Deduplicate by retaining one reading per 60 s (or averaging) before any cadence or correlation analysis. - Duplicate sensor name:
density-01appears twice (two IDs). This is the hardware replacement, not a duplicated device; see Section 3.2.
9.2 Spikes
We flag a reading when its value is at least 5× the centered 9-point rolling median (plus 0.7) and at least 10. This is a screen for single-slot outliers, not a confirmed error label.
| device | candidate spikes | max ratio | examples |
|---|---|---|---|
| goetzentuem-01 | 3 | 10.0 | 2026-08-01 08:09; 2026-10-07 20:16 |
| hagenbuchersee-02 | 7 | 22.0 | 2026-08-22 17:23 (46.2 between 1.4 and 0.7); 2026-09-19 21:29 |
| density-01 (legacy) | 4 | 10.0 | 2025-11-22 02:43 (24.0 between 7.0 and 1.0) |
| marahaus-01 | 1 | 5.1 | 2026-10-17 10:01 |
| soleo-01 | 7 | 9.3 | 2026-06-29 23:40; 2026-07-19 10:26 |
| neckarufer-01 | 6 | 11.0 | 2026-08-14 16:57; 2026-09-08 17:07 (77.0 between 11.2 and 69.3) |
| hagenbuchersee-01 | 4 | 8.0 | 2026-08-14 19:19 (292.6 between 52.5 and 65.8); 2026-10-17 10:04 |
| stadtgalerie-01 | 6 | 10.5 | 2026-08-11 10:43; 2026-08-13 11:16 |
| density-01 (new) | 1 | 16.0 | 2026-10-09 02:19 (11.2 between 0.0 and 0.0) |
There are 39 candidate spikes across the nine core devices, i.e. 0.041% of the 96,146 core readings. Some are clearly isolated single-slot outliers (46.2 between 1.4 and 0.7; 11.2 at 02:19 between zeros; 292.6 between 52.5 and 65.8). Others sit inside genuine high-traffic periods: 18–20 June shows a sustained 3–6× evening peak at several locations, and 255.5/262.5 values are surrounded by 200–250, so they are more likely real events than glitches. Without ground truth we cannot separate real surges from double counts. For the paper, report spike counts and run a winsorized/robust sensitivity analysis rather than silently deleting them.
9.3 Flatlines
- The longest run of identical consecutive readings is 20 readings (5.0 h), mostly zeros or 0.7 at night. The next longest are 17 (4.25 h, 1.4) and 16 (4.0 h, 0.0). There is no evidence of a multi-day stuck nonzero value.
- Adjacent-equal share is 7.1–29.3%, driven by zero inflation and the 0.7 grid (at low counts, 0.7 and 1.4 repeat naturally). The legacy device has the highest share (29.3%) because of its 29.8% zeros.
- Conclusion: flatlines are not a major failure mode in this export. The apparent “stuck” runs are short and occur at night/low traffic; they should not be mistaken for sensor freezes.
9.4 Dropouts
- Within the panel, seven of the eight current devices have 4–9 gaps >1 day each, with maximum gaps of 25–152 h;
density-01(new) has none and a maximum gap of 4.6 h. The long gaps are dominated by the August–September systemic outage and a few device-specific multi-day losses (e.g.marahaus152.3 h,stadtgalerie141.8 h). - The legacy
density-01has 417 gaps >1 h and 3 gaps >1 day over its 147-day life, plus 29.8% zeros; it is the least reliable series in the export. density-01(new) is the most reliable: 13 gaps >1 h, 0 gaps >1 day, max 4.6 h, and 95.2% of its cadence ceiling.
10. What these aggregates cannot measure
This section is deliberately explicit because the paper is about reliability and privacy of BLE PAX counters. The export is an aggregate product; several quantities of interest are not identifiable from it.
- MAC randomization. There are no MAC addresses, address-type flags (public vs random), RSSI, or per-detection records. The 0.7 value grid may encode a vendor correction for randomization, but the export provides no metadata to confirm this. Any statement about randomization rate from these aggregates would be speculation.
- Penetration rate. There is no independent ground-truth count (manual counts, turnstiles, sales, vehicle counts) and no denominator. The values cannot be converted to “share of visitors carrying a BLE device” or “share of devices detected.”
- Unique visitors, dwell time, repeat visits. The counters report a 15-minute aggregate per location. They do not expose device trajectories, session lengths, or cross-location matches. Cross-sensor correlation is not deduplication; it can be driven by shared city rhythm, weather, or events.
- False positives/negatives and detection radius. There is no labeled validation set. The 0.7 grid, zero inflation, and spikes are consistent with vendor processing, but the false-positive rate, missed-detection rate, and effective radius are unknown.
- Missingness mechanism. The export has no status/error codes. The August outage is visible as missing rows, but its cause (backend, gateway, network, power, firmware) cannot be determined from the CSV. Battery telemetry is not independent evidence because it disappears with density and is confounded with device identity in August.
- Unit and scale. Values are not raw counts; the 0.7 factor is undocumented. The legacy/new
density-01break proves that the unit can change across hardware/firmware. Summing or averaging across generations, or comparing absolute levels across sites, is not defensible without vendor documentation or an overlap calibration. - Causal effects. The hourly/weekday profiles and correlations are observational and confounded by season, events, holidays, weather, and the August outage. They describe co-movement, not causal impact.
- Privacy risk. The data are location-level aggregates, but low counts (0–5) at specific times and places can be sensitive. We did not attempt a k-anonymity or differencing attack; the note only assesses data quality, not disclosure risk.
Bottom line for the paper: the 0.7 grid and zero/low-count shares are proxies for vendor processing, not measurements of MAC randomization or penetration. Use them to motivate uncertainty and to rule out naïve “counts = devices” interpretations; do not present them as direct estimates of privacy-protection effectiveness.
11. Implications and recommendations
- Do not treat values as raw counts. Describe them as vendor-scaled 15-minute indices. Report the 0.7 grid and the legacy 1.0 grid in the methods; do not sum across device generations.
- Handle
density-01as a structural break. Exclude the legacy series from the panel, or model it with a separate scale/level. There is no overlap, so a conversion factor cannot be estimated; any pooling is an assumption, not a calibration. - Use 30–60-minute bins for complete-case cross-sensor analysis. At 15 minutes, 72% of slots lack at least one location; at 30/60 minutes, all-eight coverage rises to 63%/70% with little change in correlation. Report the binning choice and the all-eight coverage.
- Model missingness explicitly. Do not listwise delete 15-minute slots without reporting how many are lost. Use hierarchical/state-space models, multiple imputation, or inverse-coverage weighting; run sensitivity analyses with and without the August–September outage. Report monthly coverage alongside every estimate.
- Use robust correlation summaries. Report Spearman/log1p and daily/detrended correlations alongside raw Pearson. The raw 15-minute matrix overstates shared structure because of the common daily/weekly profile. Use block bootstrap or daily-level aggregation for uncertainty; nominal 15-minute N is not an effective sample size.
- Screen spikes and near-duplicates. Deduplicate intervals <60 s (106 of 113 are legacy), flag isolated spikes (39 candidates, 0.041% of core readings), and run winsorized sensitivity analyses. Do not silently remove the 292.6 maximum or the 46.2 outlier; report both as candidate errors/real events.
- Treat battery as a confounded covariate, not a cause. It is coarse, stale, and structurally missing with density. Within a device it does not predict daily availability; across devices it is collinear with the August outage (r = 0.84) and with device batch/gateway. If it is used at all, model it with measurement error and do not interpret the association causally.
- For privacy claims, collect the right data. To measure MAC randomization or penetration, the study needs raw scan records with hashed MACs (salted), address-type flags, RSSI, timestamps, and independent ground truth. The aggregate export cannot answer those questions.
- Pre-specify exclusions. The three episodic devices (especially the Hannover
maker-fairdeployment) and the legacydensity-01should be excluded or analyzed separately. The June 18–20 high-traffic period and the August outage should be pre-specified as events/robustness checks, not discovered post hoc. - Report the limits with the estimates. A reliability claim based on this export should carry the coverage, quantization, missingness, and unit caveats in the same table or paragraph as the estimate.
12. Outputs and reproducibility
Script: paper/research/scripts/01_data_quality.py
Command: uv run --with pandas --with matplotlib python paper/research/scripts/01_data_quality.py
Raw data: data/raw/schwarz-pax.csv (read-only; never modified)
Tables (paper/research/outputs/):
schema.csv,sensor_inventory.csvcoverage_metrics.csv,monthly_coverage.csv,monthly_diagnostics.csv,coverage_bias.csvcadence_by_sensor.csv,cadence_bins.csv,cadence_jitter.csvzero_lowcount.csv,quantization_by_device.csv,density01_scale_comparison.csvcorrelation_pearson.csv,correlation_spearman.csv,correlation_log1p.csv,correlation_detrended.csv,correlation_daily.csv,correlation_pair_n.csvaggregation_sensitivity.csv,locations_per_slot.csv,panel_15min_matrix.csv,hourly_profile.csv,weekday_profile.csv,cv_by_hour.csv,missingness_by_hour.csvbattery_summary.csv,battery_vs_coverage.csv,august_battery_coverage.csvspikes.csv,top_spikes.csv,flatlines.csv,large_gaps.csv,near_duplicate_intervals.csvintegrity_checks.csv,summary.json
Figures (paper/research/figures/):
fig_time_series.png— per-location daily seriesfig_correlation_heatmap.png— Pearson, Spearman, hour-of-week-removedfig_missingness_heatmap.png— daily slot coveragefig_distribution.png— histogram, CCDF, per-device boxplotsfig_cadence.png— interval distribution and per-device cadencefig_hourly_profile.png— hourly and weekday profilesfig_battery.png— battery time series and distributionfig_quantization.png— value-grid membership and density-01 laddersfig_density01_swap.png— legacy/newdensity-01daily means and 35.2-day gap
All figures are generated from the raw CSV by the script; no manual edits or external data were used.
Appendix A. Largest gaps (>24 h)
| device | gap start | gap end | hours |
|---|---|---|---|
| hn-bollwerksturm-01 | 2026-06-16 18:20 | 2026-08-13 16:00 | 1,389.7 |
| hn-marahaus-01 | 2026-08-19 12:25 | 2026-08-25 20:45 | 152.3 |
| hn-stadtgalerie-01 | 2026-08-20 04:46 | 2026-08-26 02:33 | 141.8 |
| hn-stadtgalerie-01 | 2026-06-12 12:11 | 2026-06-17 16:31 | 124.3 |
| hn-neckarufer-01 | 2026-06-12 12:21 | 2026-06-17 16:33 | 124.2 |
| hn-goetzentuem-01 | 2026-06-12 12:27 | 2026-06-17 16:36 | 124.1 |
| hn-marahaus-01 | 2026-06-12 12:36 | 2026-06-17 16:37 | 124.0 |
| hn-marahaus-01 | 2026-09-04 10:55 | 2026-09-09 07:50 | 116.9 |
| hn-soleo-01 | 2026-09-18 00:13 | 2026-09-20 10:18 | 58.1 |
| density-01 (legacy) | 2025-11-28 17:20 | 2025-12-01 03:59 | 58.7 |
The 12–17 June gaps are the initial deployment/test-message gaps, not operational outages; the panel starts after them. The August–September gaps are the systemic outage. The full list is in outputs/large_gaps.csv.
Appendix B. Exact cadence sequences
Legacy 863f75b0 (first readings):
2025-10-16 21:17:59.316
2025-10-16 21:35:00.506 +1021.190 s (17.020 min)
2025-10-16 21:50:42.960 + 942.454 s (15.708 min)
2025-10-16 22:05:54.445 + 911.485 s (15.191 min)
2025-10-16 22:40:43.640 +2089.195 s (34.820 min; missed ticks)
2025-10-16 22:55:56.406 + 912.766 s (15.213 min)
New f8c19730 (first readings):
2026-04-16 23:25:31.574
2026-04-16 23:41:38.142 +966.568 s (16.109 min)
2026-04-16 23:57:47.513 +969.371 s (16.156 min)
2026-04-17 00:13:56.263 +968.751 s (16.146 min)
2026-04-17 00:30:05.097 +968.833 s (16.147 min)
2026-04-17 00:46:12.591 +967.495 s (16.125 min)
The legacy series starts with startup transients and missed ticks, then settles near 911 s. The new series is stable near 968.5 s (16.14 min) with only ±2.4 s jitter. Neither is a 900 s clock.








