28 Aug 14:20 UTC

Settings

Baselines fitted
9One per metric per slice
Findings produced, 30d
31Comparisons against a fitted normal, not a fixed threshold
Suppressible by 1σ
12What a blanket widening would have cost
Oldest recompute
168hAll within cadence

Current normal, per metric per slice

Rebuffer ratio · Revolut BR · Android · GRU

B-001recomputed 2h ago · hourly
Expected
1.34%
Window
28d
Seasonality
Hour of week
Slice keys
tenant_id × platform × cdn_pop
Findings 30d
11
band 0.05%2.63%

Current11 findings in 30 days. Widening by 1σ would have silenced 4.

The busiest baseline in the system. Evening peaks in BR are strong enough that a flat mean fires nightly — the hour-of-week fit is what makes it usable.

Rebuffer ratio · Nexo Bank MX · iOS · QRO

B-002recomputed 2h ago · hourly
Expected
1.12%
Window
28d
Seasonality
Hour of week
Slice keys
tenant_id × platform × cdn_pop
Findings 30d
2
band 0.25%1.99%

Current2 findings in 30 days. Widening by 1σ would have silenced 1.

Crash-free sessions · Android TV · 3.4.0

B-003recomputed 6h ago · daily
Expected
99.30%
Window
30d
Seasonality
flat mean
Slice keys
platform × app_version
Findings 30d
6
band 99.08%99.52%

Current6 findings in 30 days. Widening by 1σ would have silenced 1.

Deliberately tight at 2σ. A crash regression that hides inside a wide band is the exact failure this tool exists to prevent.

Video start failure rate · Nexo Bank MX · iOS

B-004recomputed 4h ago · hourly
Expected
0.81%
Window
28d
Seasonality
Hour of day
Slice keys
tenant_id × platform
Findings 30d
4
band 0.39%1.23%

Current4 findings in 30 days. Widening by 1σ would have silenced 2.

Time to first frame, p95 · Android · 3.4.1

B-005recomputed 4h ago · hourly
Expected
3.24s
Window
28d
Seasonality
Hour of day
Slice keys
platform × app_version
Findings 30d
3
band 2010ms4470ms

Current3 findings in 30 days. Widening by 1σ would have silenced 2.

Activation gap · Revolut BR

B-006recomputed 168h ago · monthly
Expected
31.2%
Window
90d
Seasonality
flat mean
Slice keys
tenant_id
Findings 30d
1
band 26.4%36.0%

Current1 findings in 30 days. Widening by 1σ would have silenced 1.

Recomputed monthly against a 90-day window, so it drifts upward with the thing it is meant to detect. A fixed commercial target would be the better instrument here.

Payment failure rate · Vela Fintech AR · card

B-007recomputed 12h ago · daily
Expected
2.10%
Window
30d
Seasonality
Day of month
Slice keys
tenant_id × acquirer
Findings 30d
2
band 1.14%3.06%

Current2 findings in 30 days. Widening by 1σ would have silenced 0.

Day-of-month seasonality matters — billing runs cluster, and a flat baseline fires on every cycle boundary.

DRM licence denial rate · Altiplano PE · Android 13

B-008recomputed 6h ago · daily
Expected
0.21%
Window
30d
Seasonality
flat mean
Slice keys
tenant_id × os_version
Findings 30d
1
band 0.03%0.39%

Current1 findings in 30 days. Widening by 1σ would have silenced 1.

Join coverage · NPAW

B-009recomputed 1h ago · hourly
Expected
95.0%
Window
Seasonality
flat mean
Slice keys
source_id
Findings 30d
1

Fixed acceptance floor at 95.0%. There is no band to widen — a coverage threshold is a decision, not a distribution.

Not a fitted baseline — a fixed acceptance floor. Kept in this table so every threshold the tool compares against lives in one place.