At 2:40 on a Tuesday afternoon, the calls on the third floor break up. Sales reps hear their prospects in stutters, a screen-share freezes mid-demo, and a nurse re-dials a pharmacy three times. By 3:15 it clears on its own. When the office manager opens the carrier portal the next morning, the month reads 100.00% available. Nothing to see. The circuit was "up" the entire time. This is the gap that honest sla reporting is supposed to close, and the gap that most headline numbers are quietly designed to hide. A link that never fully drops can still be unusable for fifteen minutes at a stretch, and a service-level agreement that only counts hard outages will never admit it happened.

Why "four nines" tells you almost nothing

A 99.99% availability guarantee — "four nines" — sounds airtight. Do the arithmetic and it allows roughly 4.4 minutes of downtime per month, or about 52 minutes a year. That sounds strict until you read the fine print on what actually counts as "down." Most carrier SLAs only start the clock when the circuit is hard-failed: no signal, no route, a dead port. Degradation that still passes traffic — high latency, packets out of order, one call in twenty dropping — usually does not count against the availability number at all.

So the plant manager whose VoIP fell apart for half an hour lived through a real outage, while the SLA recorded nothing. The availability figure measures the wrong thing: whether the circuit exists, not whether it works. This is a different problem from the one we cover in our guide to reading MSP contract SLAs, which is about response and resolution commitments on a service contract. Here the subject is the physical performance of the circuit itself, and the metrics that describe it are latency, jitter, and packet loss — not a single uptime percentage.

Averages lie; percentiles tell the truth

The second place SLAs mislead is the latency number. A carrier report that says "average round-trip latency: 22 ms" describes a whole distribution by its friendliest point. Real-time applications do not fail on the average; they fail on the bad moments. If latency sits at 20 ms most of the day but spikes to 180 ms every afternoon when a neighbor on the shared segment kicks off a backup, the daily average barely moves — and yet those spikes are exactly when your calls garble.

Percentiles fix this by asking a better question: what happens at the edges, not just in the comfortable middle?

  • p95 latency. The value that 95% of your measurements came in under. It exposes the recurring bad stretches an average smooths away. If your p95 is 90 ms while your average is 22 ms, you have a link that misbehaves often enough to matter.
  • p99 latency. The 99th percentile — the worst 1% of moments. This is where the rare, severe spikes live, the ones that drop a call entirely or stall a database transaction. For voice and video, p99 sits much closer to what your users actually remember.
  • Jitter. The variation in packet arrival timing. Steady 40 ms latency is fine for a call; latency swinging between 10 ms and 90 ms is not, because the jitter buffer cannot keep up and audio breaks apart. Jitter above roughly 30 ms is where voice quality starts to suffer.
  • Packet loss. The share of packets that never arrive. Even 1% sustained loss is audible on a call and painful for a VPN; real-time media wants loss held well under half a percent. Loss is often the truest single sign that a circuit is congested rather than clean.

A performance SLA worth signing states thresholds for these four in percentile terms — for example, p95 latency under a stated ceiling, jitter under 30 ms, and packet loss under 0.1% — not a lone availability figure. If a provider will only commit to "99.99% uptime" and will not put numbers on percentiles, jitter, and loss, they are committing to the metric that is easiest to pass.

Who is holding the stopwatch?

The third question is the one buyers skip most often: who measures the numbers? Almost always it is the carrier, using probes inside their own network and aggregating over windows that flatter the result. That is not fraud; it is a conflict of interest baked into the format. A carrier measuring from their core router to their own edge can produce a beautiful graph while your office — one hop further out, on the last mile they do not instrument — sees something worse.

This is why independent, third-party monitoring matters. A system that lives at your site, sending synthetic traffic across the circuit and recording latency, jitter, and loss from where your users actually sit, produces an evidence trail the carrier did not author. When the two records disagree, the independent one is what lets you file an SLA-credit claim and make it stick — "your portal showed 100% while our probe recorded 6% packet loss for 40 minutes" is far harder to wave away than a complaint that calls sounded bad. We go deeper on how that instrumentation works in our overview of network monitoring, and continuous, site-side measurement is the core of our circuit monitoring service.

Match the SLA to the circuit type

Enforcement also depends on what you actually bought. A true dedicated circuit comes with a symmetrical, contractual performance SLA and defined credits; consumer-grade broadband comes with "best effort" and essentially no teeth. If you are weighing the two, our comparison of dedicated internet access versus broadband lays out why the SLA — not just the advertised download speed — is often the real reason to pay more. There is no point demanding p95 guarantees on a service that was never sold with them.

What a transparent monthly report contains

Once you know what to ask for, a good report is easy to recognize. Per circuit, every month, it should give you:

  • Availability with a real definition. The uptime percentage plus a plain statement of what counted as an outage, including whether severe degradation, not just a hard failure, trips the clock.
  • Latency as a distribution. Not just the average, but p95 and p99, ideally with a time-of-day view so recurring congestion windows are visible at a glance.
  • Jitter and packet loss trended over time. Daily or hourly, so a slow slide toward congestion shows up weeks before it becomes a crisis on a call.
  • The measurement source, stated plainly. Where the probes sit and who owns them, so you know whether you are reading the carrier's optimistic view or an independent one.
  • Every SLA breach, logged with timestamps. The specific windows where a threshold was missed, formatted to attach directly to a credit request — not buried inside a monthly aggregate.

A report that offers only a single green "99.99%" tile is a summary engineered to end the conversation; the version that earns trust hands you the raw shape of the month and lets you check its work.

The Bottom Line

A "four nines" promise measures whether your circuit exists, not whether it works — and an average-latency number quietly deletes the exact moments your users notice. Reading a telecom SLA honestly means insisting on percentiles, jitter, and packet loss, and on knowing who is holding the stopwatch, because independent measurement is what turns a complaint into a credited claim and a stronger position at renewal. If you would like a clear-eyed look at how your circuits actually perform versus what your carrier reports, our telecom and WAN team can stand up independent monitoring and translate the numbers into leverage. Start the conversation through our contact page or call 850-338-6503, and stop taking the uptime tile at its word.