What does the δ = 0.01 budget actually mean?

That over the entire lifetime of the monitor — not per check, per day, or per look — the probability it ever pages while your metric's true miss rate is still within tolerance is at most 1%. You can check it every second forever and the number does not degrade; that is what anytime-valid means.

How can a false-alarm probability survive unlimited peeking?

Because the statistic is a nonnegative supermartingale under the null: its expected value never grows while things are healthy, and Ville's inequality bounds the chance it EVER exceeds 1/δ by δ — a statement about the whole trajectory, not about any single look. Classical p-value monitoring has no analogous property, which is why repeated checking breaks it.

What is an e-value, concretely?

The payoff of a fair bet against the null hypothesis: it starts at 1, stays small in expectation while the null is true, and compounds when reality departs from it. An e-value of 100 means the betting evidence multiplied a hundredfold — and unlike a p-value, e-values from sequential data remain valid however and whenever you decide to look.

Coverage e-process vs e-detector — why does the product run both?

The e-detector restarts its bets against every possible changepoint, optimizing time-to-detection under a run-length budget; the coverage e-process compounds one bet from the start, optimizing the strength and auditability of the final statement under a lifetime δ cap. Fast page and certificate page — an incident typically triggers the first, and the second arrives with the evidence you keep.

All detectors

Coverage e-process

The coverage e-process watches whether your metric keeps landing inside its calibrated conformal interval, and pages only when the evidence of excess misses crosses a level with a hard cap: the probability that it ever fires while the true miss rate is still at or below α is at most δ — at every look simultaneously, for the lifetime of the monitor.

PAGE

Also known as: anytime-valid coverage certificate

This is the certificate half of ValidAnytime's page tier. Where the e-detector optimizes reaction time under a run-length budget, the coverage e-process delivers the auditable statement — an e-value you can put in a postmortem — and its quiet is the strongest quiet the product offers.

The rule, exactly

miss_t = 1{ y_t falls outside the conformal interval }
e_t = a betting e-value against “true miss rate ≤ α”
log E_t = log E_{t−1} + log e_t        (E_0 = 1)
page when E_t ≥ 1/δ

E_t is a nonnegative supermartingale under the null, so Ville's inequality caps P(E ever ≥ 1/δ) at δ. The cloud default is α = 0.1, δ = 0.01: intervals built to miss 10% of the time, and a lifetime 1% cap on ever paging while that holds.

Live playground — synthetic data, labeled as such; runs entirely in your browser

break size3σ per 16 pts

miss tolerance α0.10false-page budget δ0.01

0 pages on the 80 calm points · catches the break 52 points after onset.

Synthetic standard-normal noise (seeded, 160 uncadenced points — points, not minutes), with a labeled break at point 81 when one is enabled. The detector code running here is the same TypeScript implementation the /try bake-off runs, parity-tested against the cloud suite.

PAGE

What is — and is not — guaranteed

Its guarantee is the strongest in the product and it is unconditional on your data's distribution: the probability of ever paging while the true miss rate is at or below α is at most δ — a lifetime false-alarm budget, valid at every look, by Ville's inequality.

It is not built for speed: betting evidence compounds deliberately, so it typically pages after the e-detector on the same incident — and its budget says nothing about detection lag, only about false pages.

Page tier: e-process alarms delivered inside a stated false-alarm budget that holds no matter how often you look. See how it scores in the benchmark.

When it wins

You need an auditable alarm: every page carries the e-value that triggered it — a portable, verifiable measure of evidence you can cite in a postmortem or hand to a regulator.
False pages are expensive: δ is a lifetime cap, not a per-check rate, so a fleet of these monitors stays quiet on healthy streams for months — the property that makes always-on public verdicts defensible.
Many streams share one budget: e-values compose across monitors (online FDR), giving a fleet-wide false-discovery guarantee no per-metric threshold scheme can offer.

Where it costs you

Honestly, its failure mode is patience, not deceit: on a fast incident it is typically the last detector to fire, and if you need speed you want the e-detector or a classical warning in front of it.
A drift that never pushes the miss rate above α is, by definition, inside the tolerance you configured — it will stay silent, correctly, even if you privately hoped it would fire.
During the conformal warm-up (the first ring of calibration points) intervals are effectively infinite and misses cannot occur, so evidence only starts accumulating once the calibrator has data.

Where it comes from

The mathematics runs from Jean Ville's 1939 inequality for nonnegative supermartingales, through the game-theoretic probability of Shafer and Vovk, to the modern betting-martingale toolkit of Waudby-Smith and Ramdas. ValidAnytime's implementation bets against the conformal miss rate; this playground runs the same TypeScript port of the production path that powers /try, golden-tested against the cloud engine.

Questions engineers ask

Related detectors

In the glossary: E-process · Ville's inequality · Conformal monitoring · Online FDR control

Run it on your own history — free, in your browser.

The /try bake-off replays your metric history through this detector and the rest of the suite side by side — false alarms and catch lag, honestly graded, nothing uploaded.

Backtest your history Start free