PagerProof

Try it — then ask it what false-positive rate per year each rule will give you. It can't tell you. Those baselines are variants of mean+kσ and smoothing (Holt-Winters) applied to autocorrelated series: the independence assumption is broken from the start, so the real FP rate can be several times the theoretical one, and nobody guarantees it.

Our block bootstrap resamples blocks of your own series, preserving its autocorrelation, and calibrates under H0 of the healthy regime. That's why the FP/year we give you is a demonstrated number with an explicit condition, not a hope. And if after reading the methodology Grafana ML turns out to be enough for you — genuinely fine. Better that than static thresholds.

That's the right question, and the answer is in the report, not in our word. (1) Every threshold change in The Proof comes with a sensitivity analysis: we replay your real historical incidents against the new threshold and show you which would have fired, and with how much margin. (2) We calibrate under H0 of the healthy regime — the threshold is placed so that normal behavior doesn't cross it more than X times per year; real anomalies are, by definition, not the healthy regime.

Two honest caveats: recall can't be rigorously measured without labeled incidents, so where your history gives insufficient statistical power we mark the rule "do not touch — insufficient data" instead of guessing. And calibrated thresholds still require human review before you apply them. We're explicit about both — see the Limitations section.

You don't have to — the model is designed so trust isn't required. (1) Self-hosted: Server runs in your infra and your metrics never leave your network; if PagerProof disappears tomorrow, you keep the calibrated thresholds and everything keeps working. (2) Public methodology: the method is published and auditable — verify the math yourself or pay any statistician to do it. The closed part is the engine that automates it, not the science. (3) Fixed price, no lock-in, money-back guarantee: your maximum exposure is known and small.

One person with an auditable method is less risky than an opaque vendor with 200 employees and a "trust our AI."

They do something else. Correlation (BigPanda: −90/95% noise) groups alerts that already fired; Watchdog (−60% pager floods) filters with opaque anomaly detection. Both act after the badly calibrated threshold fires. We fix the origin: the rule firing correctly in the first place. The layers are compatible — in fact, calibrating the origin gives your correlation layer better raw material to work with.

You surely do — the question is opportunity cost. A block bootstrap under H0 done properly (block sizing, healthy-regime exclusion, multiple-rule correction) is 2–4 weeks of an engineer with a statistics background, plus maintenance. The audit costs less than one week of that same engineer, lands in 2–3 weeks, and the published methodology lets you verify it instead of rebuilding it. If you later want to internalize it, The Proof documents the path — no lock-in means that too.

The report is the format, not the product. The product is up to 27 hours/week of engineering recovered (reliamag/oxmaint) and an on-call rotation that stops burning people out. At a loaded European engineer cost (~€50/h), the audit pays for itself in under 3 weeks if it recovers just a third of that. And if the report doesn't demonstrate the improvement, you don't pay — that's the guarantee.