Metric Signals
Metric Signals is the metrics half of Intelligence: instead of a wall of dashboards you read yourself, it surfaces the metrics that moved — ranked by how far each one deviated from its own recent history — and hands each one off to the raw data with the arithmetic behind the rank on screen. It gives you the ranked signal directly, instead of a dashboard you read yourself.
It lives under Intelligence → Metric Signals. Where Critical Paths ranks request paths, Metric Signals ranks individual metric concerns; both are the “here is what to look at” surfaces that hand off to the user-driven explorers.
What you see
Section titled “What you see”The surface is a three-panel triage view, with an optional overview band across the top.
Workload overview
Section titled “Workload overview”Toggle Overview (top right) for a treemap of your services, each tile sized by request volume and coloured by statistical severity. Click a tile to filter the concern list to that service. Area and rank disagree on purpose: volume is a tiebreaker the score deliberately excludes, so a small-but-badly-deviating service still ranks above a large calm one.
What needs attention — the ranked list
Section titled “What needs attention — the ranked list”The left panel, headed What needs attention, is the concern list in fixed rank order. Each concern is a card:
- A severity pill —
LOW/MEDIUM/HIGH/CRITICALwith the numeric score beside it. Hover it for the provenance: the deviation, the confidence, and which estimator produced it. - Its rank —
#1,#2, … — the position the detector assigned, not a tier. - A title describing the concern, written by analysis.
- Its scope — namespace / service, and the operation or emitter the concern is about.
- A sparkline with an expected-range sleeve, so you can see the movement, not just the current value, and a footer with the observed value and how many standard deviations out it is.
Evidence and the Finding
Section titled “Evidence and the Finding”Selecting a concern fills the centre and right panels:
- The evidence workspace (centre) charts the observed series against its baseline and expected
range, and shows a deterministic evidence strip that spells the score out as
deviation × confidence × signal-type weight— severity is a projection of that number, not the sort key. Below it, Correlated concerns lists other concerns in the window that move with this one. - The Finding panel (right) explains why it ranks here, offers Copy for agent (the finding and its ranking arithmetic, as text to paste into your own LLM), and a Handoff preview showing exactly which query the explorer will open with.
The controls
Section titled “The controls”Three selectors in the header shape what is evaluated:
- Time window —
5 minthrough1 day— the window detection runs over. - Baseline —
7dor30d— how far back the “own history” comparison reaches. - Sensitivity — from All anomalies to Critical only, defaulting to Tenant default. Lower shows more; higher shows only the extreme. This is the recall dial; there are no per-metric thresholds to author.
How the ranking works
Section titled “How the ranking works”Every concern’s rank is a continuous score — the deviation from baseline, scaled by a confidence term and a per-signal-type weight — computed on the server against the service’s own history. atrim.ai numbers that order 1-based and never reorders it in the browser; the only rearrangement is moving telemetry with no clear service identity to the end, clearly labelled. Request volume is never a term in the score — it is only ever a tiebreaker for impact, never a reason something ranks. An LLM names the findings; it cannot change a severity or a rank, and is not given the underlying figures.
Coverage, in plain sight
Section titled “Coverage, in plain sight”The ranking is never a black box. A footer under the list shows how many series were evaluated and how many surfaced concerns, so the basis for the order is always on screen. A quiet window says so directly instead of implying more than it looked at, and while baselines are still building from your history the surface tells you that too — the signal sharpens as more data arrives.
Read-only by design
Section titled “Read-only by design”Metric Signals surfaces and explains, and holds none of your credentials. The handoff buttons only navigate to read-only explorers; Analyze is an on-demand explanation that cannot change the statistics behind a concern; and the ranking only ever uses the numbers you sent — atrim.ai never invents a figure it was not given.