ValidAnytime
  • Try it free
  • Observatory
  • Pricing
  • Docs
Sign inStart free
ValidAnytime

A guaranteed false-alarm budget across your whole fleet, valid no matter how often you look.

Made by Compiled Intelligence — a frontier AI lab working on quantitative finance from first principles; ValidAnytime is the monitoring we built for our own model fleets, productized.

Product

  • How it works
  • Pricing
  • Observatory
  • Try it free
  • Dashboard

Solutions

  • Who it's for
  • For LLM engineers
  • For MLOps engineers
  • For SRE & platform
  • LLM eval regression
  • Model drift detection
  • Alert fatigue

Developers

  • Docs
  • Quickstart
  • Python SDK
  • API reference

Learn

  • Benchmark
  • Detector guides
  • Glossary
  • Compare
  • Why dashboards lie
  • Blog
  • Anytime-valid 101
  • Winter Storm Uri case study
  • Incident ledger
  • Observatory methodology

Company

  • Contact
  • Privacy
  • Terms

Observatory alarms & product updates, by email

Rare emails only. We store your address and the tag “footer” — nothing else. Unsubscribe anytime.

© 2026 ValidAnytime. All rights reserved.

Every alarm ships with its guarantee_tag and theorem_ref.

Home/Solutions/Use case · Alert fatigue

Use case · Alert fatigue

Alert fatigue is a math problem. Here's the math that fixes it.

A fixed threshold checked every minute will eventually fire on nothing — so teams mute the channel and miss the real one. That's not carelessness; it's arithmetic. ValidAnytime replaces the fixed line with an anytime-valid detector, so looking more often makes you safer, not noisier — one alarm per real change, with a false-alarm budget across the whole fleet.

Start freeSee the peeking problem

Free to start · no credit card

The alert everyone learned to ignore.

Count the alerts your team muted this month. Each one was rational: it had cried wolf so many times that ignoring it was the correct move. That's the trap — every mute is a small, sensible decision, and together they train the whole team to sleep through the one page that mattered. The runbook that says “wait for the second alert” is that training, written down.

Why “just tune the thresholds” never ends.

Tighten the threshold and you catch more real events — and far more false ones. Loosen it and the noise stops, and so do the early warnings. No setting fixes both, because a fixed-sample test is being checked continuously: check a fixed 5%-error test 5 times and the real false-positive rate is already ~23%; at 20 looks it passes 64%. Add monitors and it compounds, with no global bound on false pages.

  • Muted channels and “ignore the first alert” runbooks.
  • Thresholds you re-tune every quarter and never get right.
  • Every new service multiplies the noise.
  • No way to say what fraction of your pages are false.

How it works for you

  1. 1

    Stream any metric

    Latency, error rate, a model score, a KPI — send the number through one HTTP call or the SDK. Any numeric metric via the API.

  2. 2

    Backtest before it pages anyone

    Replay your history: the config must stay quiet on your normal weeks and fire on the incident you already know about before it can wake a human.

  3. 3

    One alarm per real change — fleet-wide budget

    E-processes accumulate evidence instead of re-rolling the dice; online FDR bounds false discoveries across every stream. The pages you get are worth acting on.

Watch the noise stop.

Drag the slider: the fixed threshold piles up false alarms the more often you check; the evidence line stays flat and fires once. The same engine runs 24/7 on public AI-vendor signals in our Observatory — with its false-alarm budget stated on the page.

Same stream, two ways of watching it
Watching:
now: —
injected drift begins · day 84LLM answer-quality scorefixed threshold ±1.9σValidAnytime — evidence of a real changealarm level
hourly

Fixed threshold — ≈ 7 false alarms

The more often you look, the more it cries wolf on stable noise.

ValidAnytime — 0 alert, 0 false

Watching — no evidence of a real change yet.

20 parallel worlds— 20 simulated healthy streams; almost none ever cross.

Roll 20 genuinely random healthy streams and watch nearly all of them stay quiet — the false-alarm bound, re-runnable yourself.

Illustrative run on a synthetic stream, scored by the same anytime-valid detector we run on your data. The metric labels above are just framing — the numbers and the detector’s behavior are identical. In onboarding, ValidAnytime replays your history and shows whether — and on which day — it would have caught your regression.

p99 latency — step + creep

Synthetic, illustrative — runs in your browser, nothing uploaded.

Watch it get caught

E-process certificate — every alarm ships one

warned_at
hour 68 — warning tier, model-calibrated, unbudgeted
paged_at
hour 78 — inside the stated false-alarm budget
e_value
20.2
fleet_review
confirmed discovery — online FDR control (e-LOND)
guarantee
average_run_length_e_detector / anytime_valid_under_conditional_mean_null

From the seeded demo — synthetic stream, real engine. Load it in the live dashboard →

Not another dashboard. The trustworthy alarm on top of one.

Keep your APM and dashboards — they're great at showing what happened, and we don't claim to replace them or to integrate with them yet. ValidAnytime fixes the one thing fixed thresholds do badly: the alarm.

  • vs Threshold dashboards (Datadog / Grafana-style)

Questions, answered

Keep reading

  • Alerts that don't cry wolf, for SRE teams
  • ML model monitoring without false alarms
  • LLM monitoring that doesn't cry wolf
  • The peeking problem
  • False discovery rate
  • Online FDR control
  • Your monitoring dashboard is lying to you
  • Static threshold — the rule behind most dashboard alerts
  • The honest drift-detector benchmark

See whether — and on which day — it would have caught your last regression.

Point ValidAnytime at one stream and it replays your own history — free, no credit card.

Free to start · no credit card · we’ll only email you about your account.