Interactive demo · Staged alerts, real Jev judgments. Routing and review logic run in your browser. No notifications are sent.

How this demo works

The alerts are the Docker demo's staged incident. The model's answers are replayed from the real jev-1.13.0 run on 2026-09-23 (no model is called in your browser). Everything else is the real pipeline: routing, the dedup graph, the review clock, shadow mode and evaluation. Sending pages to PagerDuty, Slack or Jira isn't built yet.

Triage timing

2.07 s

8 decisions

8alerts

3 Page

2 Review

2 Ticket

1 Quiet

Latency per decision

p50228 ms

p1178 ms

p95753 ms

Demo clock

14:04 UTC

15× speed

Your name

Reviews waiting for an ack 2

Ack means someone is on it: the review closes and won't escalate to a page. It doesn't resolve the incident; the alert stays open until your monitoring clears it. If nobody acks before the deadline, the review escalates to a page.

PaymentServiceErrors: payment authorization failures 18%paymentcomputepages at 14:18 UTC unless acked
14m 45s left
TlsCertExpiringSoon: TLS certificate for api.example.com expires in 25 daystls-certnetworkpages at 14:18 UTC unless acked
14m 45s left

3 paged; 2 in review.

3 ticketed, logged or dropped. Of 2 sent to review, 2 still waiting.

Page#1OrdersDbPoolExhausted: connection pool 100% used (500/500) on the orders-db primary→Paged753 ms

page 1.00

quiet 0.04

database

Quiet#2StagingDiskFull: disk 97% full on staging CI runner→Logged

triage-rotation env=staging: non-production never pages (rule, no model call)

Ticket#3SearchApiHighMemory: memory 88% of limit on search-api→Ticketed234 ms

page 0.07

act 0.31

quiet 0.62

compute probably noise, but P(actionable) > 0.05: ticket, not drop

Page#4CheckoutApiErrorRate: 5xx rate 31% for 5m→Paged now200 ms

page 1.00

quiet 0.03

database

Review#5PaymentServiceErrors: payment authorization failures 18%→Sent to review221 ms

page 1.00

quiet 0.06

compute duplicate_of checkout (0.87); incident root checkout

Review#6TlsCertExpiringSoon: TLS certificate for api.example.com expires in 25 days→Sent to review257 ms

page 0.47

act 0.49

quiet 0.04

network unsure whether to page (0.2 < P(page) < 0.8): a human decides within 15m or it pages

Ticket#7ReportJobSlow: nightly revenue report finished 12 minutes later than usual→Ticketed228 ms

act 0.22

quiet 0.78

compute probably noise, but P(actionable) > 0.05: ticket, not drop

Page#8HomepageLatencyHigh: p99 latency 2.4s against a 400ms SLO→Paged178 ms

page 1.00

quiet 0.04

compute

Inspector

Auto-cycling

Select an alert from the feed to see its analysis.

Against the labels

2 labeled alerts is a smoke test. Trust these numbers, and the calibration especially, only after replaying a few hundred labeled alerts.

Outcomes

What happened	jev-oncall	Your current routing
Needed a human, reached no one	0	0
Should have paged, got a ticket	0	0
Should have paged, got a review	0	0
Paged when it shouldn't have	0	1
Paged twice for one incident	0	0
Reached the wrong team	0	n/a
Linked to the wrong incident	0	n/a
Review that wasn't needed	0	0
Ticket for noise	1	0
Pages sent, all 8 alerts	3	5
Reviews sent, all 8 alerts	2	0

Rows above the line count the 2 labeled alerts; pages and reviews count every alert, as decided on arrival. Your current routing means configured severity only: critical pages, warning tickets, info logs, in every environment. jev-oncall also never pages for non-production alerts, which is part of the difference.

Agreement with labels

Question	Matched	95% interval
Actionable	2 of 2	34% to 100%
Page or not	2 of 2	34% to 100%
Severity	1 of 2	9% to 91%
Owner	2 of 2	34% to 100%
Cause	2 of 2	34% to 100%

Calibration of P(page)

P(page)	Alerts	Predicted	Paged by label
0.0 to 0.2	1	0.00	0.00
0.8 to 1.0	1	1.00	1.00

Brier score 0.000, calibration error 0.000. Lower is better for both.

About this run

Model: jev-1.13.0
Alerts: 8: 7 judged by Jev, 1 by rule, 0 by fallback
Cost: $0.00000 for 0 input tokens
Latency per call: 228 ms median, 753 ms p95
Policy: Page at 0.80, review above 0.20, drop at 0.05 actionable, link at 0.70
Safety checks: All invariants held

Compared with your current routing

Since Sep 24, 2026, 14:03 UTC: 8 alerts. Your current routing paged 5; jev-oncall paged 3 on arrival. 2 alerts labeled so far. Select any alert to label it in the inspector.

Your current routing means configured severity only: critical pages, warning tickets, info logs, in every environment. jev-oncall also never pages for non-production alerts, which is part of the difference. The evaluation below compares the same two.

3Same as your current routing
1Would page, your routing didn't
1Would ask for a review, your routing didn't page
1Your routing paged, would ask for a review instead
2Your routing paged, would not page

Latest differences	Your routing → jev-oncall
StagingDiskFull: disk 97% full on staging CI runnerenv=staging: non-production never pages (rule, no model call)	PAGE → LOG
PaymentServiceErrors: payment authorization failures 18%P(page)=1.00 P(actionable)=0.94 team=compute@0.99	PAGE → REVIEW
TlsCertExpiringSoon: TLS certificate for api.example.com expires in 25 daysP(page)=0.47 P(actionable)=0.96 team=network@1.00	TICKET → REVIEW
ReportJobSlow: nightly revenue report finished 12 minutes later than usualP(page)=0.00 P(actionable)=0.22 team=compute@1.00	PAGE → TICKET
HomepageLatencyHigh: p99 latency 2.4s against a 400ms SLOP(page)=1.00 P(actionable)=0.96 team=compute@0.97	TICKET → PAGE

#	Decision	P(page)	ms
#1	page	1.00	753
#2	quiet	-	-
#3	ticket	0.07	234
#4	page	1.00	200
#5	review	1.00	221
#6	review	0.47	257
#7	ticket	0.00	228
#8	page	1.00	178

1.00

Paged

orders-db · orders-db · 14:03 UTC

1page1.00

2SEV21.00

3database1.00

4actionable0.96

5SEV10.00

6SEV30.00

Severity: SEV1 0.00, SEV2 1.00, SEV3 0.00, SEV4 0.00
Owner: database 1.00
Cause: none at 1.00
Candidates: search-mem, checkout, payment, tls-cert, report-job, homepage
Call: unknown input tokens, 753 ms
Configured: critical
Labeled: actionable, SEV1, database

SEVERITY

0.00

Logged

staging-disk · ci-runner · 14:03 UTC · staging

Configured: critical

ROUTING REASONS

env=staging: non-production never pages (rule, no model call)

0.07

Ticketed

search-mem · search-api · 14:03 UTC

1compute1.00

2SEV30.93

3actionable0.38

4page0.07

5SEV20.07

6SEV10.00

Severity: SEV1 0.00, SEV2 0.07, SEV3 0.93, SEV4 0.00
Owner: compute 1.00
Cause: none at 1.00
Candidates: checkout, orders-db, payment, tls-cert, report-job, homepage
Call: unknown input tokens, 234 ms
Configured: warning

SEVERITY

ROUTING REASONS

probably noise, but P(actionable) > 0.05: ticket, not drop

1.00

Paged now

checkout · checkout-api · 14:03 UTC

1page1.00

2SEV11.00

3database0.99

4actionable0.97

5SEV20.00

6SEV30.00

Severity: SEV1 1.00, SEV2 0.00, SEV3 0.00, SEV4 0.00
Owner: database 0.99
Cause: none at 1.00
Candidates: orders-db
Call: unknown input tokens, 200 ms
Configured: critical

SEVERITY

1.00

Sent to review

payment · payment-service · 14:03 UTC

1page1.00

2SEV21.00

3compute0.99

4actionable0.94

5SEV10.00

6SEV30.00

Policy override: On its own, its P(page) says page. It joined incident checkout, and database owns checkout, so compute gets a review instead of a page: a person on the right team decides, and nobody is paged twice for one incident.
Review: Waiting for an ack since 14:03 UTC. If nobody acks by 14:18 UTC, it pages.
Severity: SEV1 0.00, SEV2 1.00, SEV3 0.00, SEV4 0.00
Owner: compute 0.99
Cause: checkout at 0.87
Candidates: checkout, orders-db
Call: unknown input tokens, 221 ms
Configured: critical

SEVERITY

ROUTING REASONS

duplicate_of checkout (0.87); incident root checkout
root is owned by database, so compute gets a REVIEW, not silence

0.47

Sent to review

tls-cert · edge · 14:03 UTC

1network1.00

2actionable0.96

3SEV30.53

4page0.47

5SEV20.47

6SEV10.00

Review: Waiting for an ack since 14:03 UTC. If nobody acks by 14:18 UTC, it pages.
Severity: SEV1 0.00, SEV2 0.47, SEV3 0.53, SEV4 0.00
Owner: network 1.00
Cause: none at 1.00
Candidates: payment, report-job, checkout, homepage, search-mem, orders-db
Call: unknown input tokens, 257 ms
Configured: warning

SEVERITY

ROUTING REASONS

unsure whether to page (0.2 < P(page) < 0.8): a human decides within 15m or it pages

0.00

Ticketed

report-job · report-batch · 14:03 UTC

1SEV41.00

2compute1.00

3actionable0.22

4page0.00

5SEV10.00

6SEV20.00

Severity: SEV1 0.00, SEV2 0.00, SEV3 0.00, SEV4 1.00
Owner: compute 1.00
Cause: none at 1.00
Candidates: homepage, tls-cert, payment, checkout, search-mem, orders-db
Call: unknown input tokens, 228 ms
Configured: critical
Labeled: not actionable, SEV4, compute

SEVERITY

ROUTING REASONS

probably noise, but P(actionable) > 0.05: ticket, not drop

1.00

Paged

homepage · homepage · 14:03 UTC

1page1.00

2SEV21.00

3compute0.97

4actionable0.96

5SEV10.00

6SEV30.00

Severity: SEV1 0.00, SEV2 1.00, SEV3 0.00, SEV4 0.00
Owner: compute 0.97
Cause: none at 1.00
Candidates: report-job, tls-cert, payment, checkout, search-mem, orders-db
Call: unknown input tokens, 178 ms
Configured: warning

SEVERITY

Reviews waiting for an ack 2

3 paged; 2 in review.

Inspector ⏸

Against the labels

Outcomes

Agreement with labels

Calibration of P(page)

About this run

Compared with your current routing

Inspector