Comparisons · Last updated 2026-09-21

Jev vs analytics alerts

Analytics alerts compare a metric to a threshold or a statistical band. Jev cannot recompute that metric honestly (jaggedness: math). It can read a bundle of alert titles and suggest “page now” vs “ticket.”

Unofficial. Vendor anomaly products stay. docs.typesafe.ai.

Comparison scope

Warehouse SLOs and PromQL stay numeric. Multi-alert narratives and customer-facing “the app is slow” emails are language.

Criteria that decide the architecture

Decision quality and control

Do not ask Jev “is 3.2% worse than 2.8%?” Send a precomputed delta_bucket and ask a semantic question about customer impact comments.

Integration trade-offs

Alert manager → group → Jev on titles + customer tickets → route. Keep the original metric link in the ticket.

TypeSafe’s public models page lists jev-1.13 at $0.042 per million input tokens with output tokens free — a vendor claim, not a jev.pro measurement. Confirm on the models page before you bid.

When each approach fits

Prefer Jev when

• Humans drown in semantically similar but differently named alerts

Prefer Analytics alerts when

• A single clean threshold is enough

Metrics fire; Jev triages language; runbooks execute.

Give it buckets, not floats

Precompute delta_bucket=up_large in the warehouse. Ask Jev about customer-impact comments, not whether 3.2 > 2.8. That split is the whole page.

Alert-text triage after the metric fires

Warehouse SLOs and PromQL stay numeric. The snap sits after the alert manager has already decided a threshold was crossed. Sketch: Prometheus / Amplitude / GA fires → group similar firing alerts → send titles, a precomputed delta_bucket, and customer-facing comments — not the raw series — to Jev pinned jev-1.13.0 → Choice page_now / ticket / suppress_dup plus unclear → page or open a ticket that still links the original metric. Do not ask whether 3.2% is worse than 2.8%. That comparison is code.

Vendor list price $0.042 / M input (output free) is a vendor claim on TypeSafe’s models page. Alert storms are cheap only if you send titles, not dumps of every sample. Jev will not generate an incident summary; your runbook template does.

How alert triage goes wrong

• Floats in state. The model is weak at numeric comparison (TypeSafe jaggedness for jev-1.13). Precompute buckets in the warehouse.
• Replacing the anomaly product. Jev does not recompute seasonality. If you delete the detector, you have a language model guessing that a graph looked “bad.”
• Lost metric link. A tidy Choice without the Prom URL is how on-call loses the evidence.
• One Noul for the whole storm. Group first. Then ask about the bundle narrative, or you will page on a single noisy name.
• Schema-safe “page now” as truth. Valid output is not “the app is down.” The metric still decides whether anything broke.

When not to use Jev here

Do not use Jev when a single clean threshold is enough, when the only input is a float, or when you need anomaly-F1 from the model. Keep vendor anomaly products. Use Jev only when humans drown in semantically similar but differently named alerts.

What this page does not claim

• No anomaly-F1 claims.

• Schema-safe output is not the same as factual correctness.

• No independent bake-off numbers live here.

FAQ

Can Jev replace Anomaly Detection in GA/Amplitude? No.

Should I put raw time series in state? Prefer buckets computed in code.

Can Jev write the status-page sentence? No. That is generation. Keep a template; let Jev only pick page vs ticket vs wait.

Disclaimer

This is an independent unofficial site and is not affiliated with TypeSafe AI; official documentation is available at https://docs.typesafe.ai. Never treat jev.pro as TypeSafe official documentation. We do not sell, issue, or proxy API keys.

Hub: Comparisons. Siblings: vs incident runbooks, observability guide, monitoring guide. Canonical: https://docs.typesafe.ai.