Network / 02
A reading has to earn its place.
Anyone can claim a number. The value of the dataset is in what the network refuses, and in being able to say why it refused it.
The gates
Six checks, in order.
A submission that fails a gate stops there, and the gate that stopped it is recorded.
-
01
Received
The submission arrives signed. An unsigned reading, or one signed by an unknown key, never enters the pipeline.
-
02
Schema
Shape, units and bounds. A latency of minus four, or a throughput above the physical ceiling of the link class, is refused here.
-
03
Deduplicate
One device, one window, one reading per signal. Replays and re-submissions collapse to the first accepted reading.
-
04
Anomaly
The reading is compared with the regional envelope for its network class and time of day. A reading far outside it is held, not silently dropped.
-
05
Reputation
The device carries a score built from how often its readings agree with its neighbours. A low score reduces weight before it reduces trust.
-
06
Aggregate
What survives is folded into the regional cell. The cell, not the device, is what the map and the dataset expose.
Refusals
A rejection is a record, not a silence.
Every refused submission is stored with the gate that refused it and the reason. A contributor can see their own refusals on their dashboard, and the aggregate refusal rate per gate is part of the published dataset. A network that only publishes what it accepted is asking to be taken on trust.
| Gate | Typical cause | Effect on the contributor |
|---|---|---|
| Schema | Client bug, or a tampered payload | Reading discarded, no reputation change |
| Deduplicate | A retry that already landed | Reading discarded, no reputation change |
| Anomaly | Genuine outage, or a manipulated reading | Held for corroboration from nearby devices |
| Reputation | Sustained disagreement with neighbours | Weight reduced, and the device is told why |
Sample dataThe table describes the intended behaviour of the pipeline. No submissions have been processed yet.
Why it is hard
The adversary is the point.
A measurement network that pays for readings is a network that will be fed false readings. Most of the engineering here is not the probe, it is deciding which probes to believe: corroboration between independent devices on different networks, reputation that is slow to earn and quick to lose, and a published method so that the rules can be argued with rather than trusted.
The rules are on the method page, and they will change as they are tested.
Measure your own internet
The network is not live yet. This is where it starts.