A case of contrasting signals · Observed State Index
Summary
Between the afternoon of 2 September 2026 and the morning of the 3rd, the system flagged nine countries as out of range at the same time in its internet section: Indonesia, Thailand, Japan, Singapore, Australia, Vietnam, Mexico, the Philippines and Hong Kong. The flag was correct, because the signal the system watches had left its range.
But that signal measures a single thing: how many networks answer a probe. Read on its own, it said something that did not happen. It was checked against three other measurements of a different nature, for the same countries and the same hours, and none of them moved. The networks did not disconnect: they stopped answering.
The lesson is not about this event but about what a single signal can claim on its own: a signal that crosses its threshold is a fact about that signal, not yet about the world.
What the system flagged
For the internet, the system uses the ping-slash24 signal from IODA, the connectivity measurement project at Georgia Tech. IODA sends probes to blocks of addresses in each country and counts how many respond. The system compares that number with the previous 90 days for the same country, and marks it «out of range» when it moves five standard deviations away or changes by more than 50%.
In IODA's public series, the drop begins in the same ten-minute interval, at 13:40 UTC on 2 September, in eight of the nine countries; in Hong Kong, ten minutes later. It ends between 05:30 and 09:40 UTC on the 3rd. In the system's run at 09:22 UTC on 3 September, all nine were out of range.
Why it looked real
Everything pointed to an event. The drops were large: at their worst interval, Indonesia lost more than half of the networks that responded, and Japan more than a third. Everything lined up: the countries formed a coherent geographic group, and the drop began at the same moment in all of them.
IODA's own detector flagged it. For Japan alone it issued seven critical alerts between 14:20 UTC on the 2nd and 09:00 on the 3rd. And a second copy of the system, running in parallel during a migration, flagged it too; that is how it was recorded in the internal documentation — as a confirmation.
It was not one. Both copies read the same IODA API. They were not two witnesses, but the same witness telling the same thing twice.
The contrast
Besides the probes, IODA publishes other signals per country, and three of them measure different things. BGP counts how many address blocks are still announced in the internet's routing system: a network that disconnects stops being announced. Traffic to Google measures how many real people reach Google's services from that country. The network telescope picks up the background traffic that switched-on machines send, unintentionally, to a range of addresses nobody uses.
When a country loses the internet, all four fall together. Each one was compared over the event window, from 12:00 UTC on 2 September to the same time on the 3rd, against the same window the day before:
| Country | Probes (worst interval) | BGP | Traffic to Google |
|---|---|---|---|
| Indonesia | −52.3% | 0.0% | −1.1% |
| Thailand | −40.0% | 0.0% | −2.1% |
| Japan | −35.7% | 0.0% | +0.6% |
| Singapore | −35.6% | 0.0% | +14.3% |
| Australia | −24.2% | 0.0% | −0.5% |
| Vietnam | −21.0% | 0.0% | +1.5% |
| Mexico | −15.3% | −0.1% | +2.4% |
| Philippines | −11.9% | 0.0% | −0.7% |
| Hong Kong | −11.3% | +0.1% | +1.8% |
Routes did not move by even a tenth of a percent in any country. Traffic to Google repeated its daily curve, and in Singapore it was even higher than the day before. The telescope, a noisy signal, stayed between −7% and +3%, with no drop comparable to that of the probes. Aviation in the region's corridors stayed at its usual level, which makes a physical event unlikely.
What it means
During those hours, the networks of those nine countries stayed announced, kept serving their users and stayed switched on. The only thing that fell, for about twenty hours, was the share of networks answering IODA's probes.
That is a fact about the measurement, not about the connectivity of those countries. Anyone reading «internet out of range in Japan» would understand «Japan is having internet problems», and that did not happen.
What cannot be known
The cause. With these data it is not possible to tell whether there was a problem on the path between the probes and that region, filtering of the responses, or a fault in the probes themselves. This document does not attribute the event to IODA or to anyone else. IODA's own method combines several signals precisely to tell cases like this apart.
What was found at home
The system stores BGP and the telescope as context signals alongside the probes. BGP was exactly the signal that would have refuted the outage that same day. And it was not being stored. In the run at 09:22 UTC on 3 September, BGP was empty for all 52 countries.
It was not a code error but a clock one. The process asks IODA for the hour that has just closed and was keeping the last interval of each signal. Twelve minutes after the hour closes, the probes are already published and BGP almost never is, so BGP was being stored empty. Nothing gave it away: an empty context value is a valid state, and no alarm was watching it.
It was fixed on 13 September. BGP and the telescope now keep their last interval that has data; the probes do not change. The first run after the fix was at 05:10 UTC on 14 September, at the same minute that used to leave it empty, and BGP arrived for all 52 countries. The earlier days cannot be recovered.
What this case shows
That a signal crossing its threshold is a fact about that signal. To claim something about the world, it takes another signal, of a different nature, saying the same thing.
That confirming with the same source read twice is not confirming, even when it looks as if two systems agree.
That a context signal is only useful if someone checks that it is being stored. This one was in the design, in the code and in the files, and it was empty.
Methodological note
All the figures in the contrast come from IODA's public API (api.ioda.inetintel.cc.gatech.edu), queried per country for 1 to 4 September 2026. Anyone can reproduce them. «Worst interval» is the lowest ten-minute interval against the median of the reference window; BGP and Google are averages over the whole window.
Malaysia also dropped (−23.5% at its worst interval) without reaching its threshold, because each country is measured against its own variability. Poland had a brief dip that afternoon that does not follow the pattern, and was left out. The aviation check covers three of the four time bands of the event: the one at 02:30 UTC on 3 September is not in the system's data. Cloudflare Radar, whose annotations of verified incidents the system uses as context, recorded no outage in those countries on those days.
← Back to the index