Docs
The method.
A reading is only worth something if the way it was taken can be checked. This is that description, in the state it is currently in.
1. What is measured
Six signals per run: latency, packet loss, jitter, throughput, DNS resolution time and stability. Each is measured against a regional reference endpoint, in a fixed order, under a hard time budget. Latency, jitter and throughput are recorded as distributions, not single figures, because a median and a 95th percentile describe a connection and a single number does not.
2. How a run is bounded
- A run has a fixed probe count and a fixed time budget. It stops when either is reached.
- Throughput uses a capped transfer, so a run cannot cost a meaningful amount of data.
- A run is skipped, not queued, if the line is already busy, so the measurement does not become the congestion.
- Runs are scheduled by the contributor, and the schedule is part of the submitted record.
3. What makes a reading acceptable
A submission must satisfy all of the following. Failing any one stops it at that gate, and the gate is recorded.
- Signed by a known device key, over the full record including the window.
- Within schema and physical bounds for its declared network class.
- Unique for that device, window and signal.
- Inside the regional envelope for its class and time of day, or corroborated by independent devices if outside it.
- From a device whose reputation is above the weighting floor.
4. What is published
Regional aggregates only. A published record names a cell, a network class and a closed window, and carries the distribution, the device count and the refusal rate. Device-level records are never published and never sold. The refusal rate is published alongside the readings, because a network that only publishes what it accepted is asking to be taken on trust.
5. Known weaknesses
- Contributors are self-selected, so coverage is biased toward better connections. Bounties reduce this and do not remove it.
- Corroboration works where device density is high and is weakest exactly where coverage is thinnest.
- An edge probe cannot attribute a bad reading to a specific hop. Attribution beyond the cell is inference.
- Throughput measurement consumes the capacity it measures.
6. Status
This document describes a design, not a running system. No reading has been taken, no gate has refused anything, and no aggregate has been published. Version and change history start when the first client ships.
Tell us where this is wrong
The method is published so it can be attacked before anything is built on it.