# Measurement caveats

- First time bucket: 49.8% of rows and 79.4% of volume. The ensuing decline cannot be presented as real network demand without understanding timestamp fabrication.
- 81.5% of subscribers occur in multiple province labels within a timestamp. Sixteen station IDs also span more than one province. The representative attributes the multi-province subscriber behaviour to location randomization; do not interpret this as movement or a network error. The station-to-province mapping remains unverified.
- Zero exact duplicate rows, but even a five-field candidate row key repeats. The actual unit of observation is unresolved.
- HTTP success-rate values exceed 1 in 26,119 rows (41.4% of non-null values). The field's scale or generation needs clarification.
- HTTP response time is zero in 94.4% of rows; 99.8% of these rows have missing HTTP success rate. Zero may mean no applicable measurement, not instant response.
- Video+audio exceeds stated total volume in 6,428 rows; tethering exceeds it in 21,818. Do not enforce subset accounting until definitions are confirmed.
- Missing throughput and RTT vary; comparisons need observed counts and should not replace missing values with zero.

See [unresolved questions](../open-questions.md) and [analysis provenance](provenance.md).