Is it Wi-Fi,
your ISP
or the Internet?

octomon is not an observability platform, and it doesn't want weeks of your network data. It answers one question: how is my connection right now, and is that normal for where I'm sitting?

Each target gets an ICMP probe every second. The w window (30s to 15m) sets how many of those outcomes the table summarises, and every cell grades itself over that window: the name and last columns display the current condition while avg, p95, max, jitter (RFC 3550 interarrival) and loss carry the history. A recovered path visibly greens from the left as the incident drains out, with a ↓ on loss that is history rather than happening.

Underneath, each network is fingerprinted on SSID plus gateway MAC, so DHCP renewals and mesh roams keep one identity and the same LAN over copper and over radio count as two different normals. What normal means is learned as slow EWMAs (exponentially weighted moving average) folded once per clean minute: latency to your gateway and to the best remote address (typical and p95), resolver response time, Wi-Fi signal strength, and crucially how much ICMP loss is usual there. A minute tainted by an active network issue never teaches the baseline; a degraded state that stands long enough is reclassified as that location's weather and does. Comparisons stay silent until five minutes have been learned, and the baseline persists across runs, so a session that starts broken is still judged against what this network can (or cannot) do.

The grading is relative with absolute floors: latency within 1.5x of the path's floor is fine and beyond 3x is bad, loss within 1.5x of the learned normal is fine, with office-LAN absolutes as the fallback when nothing has been learned yet. That's what keeps plane Wi-Fi honest. It drops 20 to 40% of ICMP as permanent weather (control-plane deprioritisation, not path failure) while TCP still flows, which octomon proves by making small web requests, over both IPv4 and IPv6 (when available), to confirm actual traffic still gets through. So instead of a wall of red and a false outage, one line: ▲ connection degraded but usable, web traffic still getting through.

Hotel Wi-Fi is the mirror case: fine at noon, congested every evening. While a latency problem is active, the latency numbers are left out of the learning, so the 9pm inflation never redefines what normal means, and the analysis can say gateway 41ms vs ~9ms normal here and name the network as the culprit. Findings themselves pass hysteresis (raise on 4 of the last 6 evaluations, clear only after a quiet stretch), so one lost packet never flaps the verdict. And the past isn't the live table's job: the events timeline e and each location's incident history keep what happened an hour ago, with timestamps and durations, so the windows can stay short and honest.