Glossary

Telemetry Nerd uses one precise word for each of these ideas, everywhere: in panel captions, in the skills Claude follows, in the specs. If you see a different word for the same thing in an older doc or issue, that’s a bug in that text, not a second term. The full reference, including the exact code identifiers behind each one, is docs/glossary.md in the repository; this page is the subset that matters when you’re reading a panel rather than writing code.

Time and resolution

• Series interval — how often the underlying data actually has a point: the scrape interval for a pulled (Prometheus) metric, the push or export interval otherwise. A property of the data itself, not something you choose.
• Query step — how far apart the points you asked for are. Each one covers a bucket of time ending at that point.
• Query bucket — one fetched point: the time interval it covers, plus its average, min, max and sample count.
• Display bucket — after zooming out, several query buckets get merged into one drawn point. The merge always keeps the min and max, so a spike doesn’t disappear (see “envelope” below).
• Time range — the start..end a query, dataset or panel covers. Never just “window” — that word is reserved for the next two terms.
• Query window — the [5m] in something like rate(x[5m]): how far back each point’s calculation looks, not the overall time range.
• Reference time range — the earlier period a comparison or verdict is judged against (the previous day, the same hour last week). Stated explicitly, chosen before the result is seen (principle 14).

A step much shorter than the series interval invents detail between real samples; a step very close to the series interval makes real samples spill across buckets unevenly. Neither is a bug in the chart — it’s what you asked for.

What you see on a panel

• Envelope (min–max band) — the shaded band behind a line, running from each bucket’s minimum to its maximum. It’s drawn so a short spike still shows its true height even after rolling up to a wide time range, instead of being averaged away.
• LOD (level of detail) — the server-side merge that produces the envelope: about one display bucket per pixel, never losing the extremes.
• Coverage rug — the strip under a time panel marking which buckets are missing, partial or untrusted, each drawn differently. Gaps are shown, never bridged with a straight line (principle 11).
• Settling — the newest few buckets, which the source may still revise (late-arriving samples, backfill). Not an error — just not final yet.
• Operating profile — what a signal normally looks like over a long history, broken down by hour of the week.
• Normal band — the per-series band drawn from the operating profile for the current hour, so an unusual value stands out without anyone setting a manual threshold.
• Ghost — the same time range from one week earlier, drawn faint and dashed for a quick visual comparison.
• Indexed view — the Y axis redrawn as a ratio to a baseline (the series’ own mean, the previous window, last week) on a log scale, so you’re looking at relative change rather than absolute units.

Model views

• Little’s law window — the sub-range Claude averages over when checking L = λW for a service (concurrency, throughput, latency); shorter than the full time range by default.
• Point model / cautious model — a test run under two sets of assumptions: the point model is the plain estimate (optimistic, shown as context), the cautious model allows for what could make the result look stronger than it is (autocorrelation, clustered events, heavy tails). A label such as “special cause” rests on the cautious model; when only the point model sees something, it is reported as undetermined (principle 16).

See the rationale page for why a model-view verdict always ships with the measured discrepancy, not just a label, and the architecture page for how the pieces that compute these views fit together.