Synthetic demo · no Descript data · metric-SEV for model-quality

Metric-SEV console for model-quality regressions

Point it at a stream of model-quality metrics, transcription word-error-rate by slice, voice-clone fidelity, an AI-editor action score, and it shows, verdict-first, which metrics would have paged someone and which drifted quietly to the next retro. The console is the argument: it makes "your quality bar isn't on an on-call" visible in one screen. Press a scenario; the verdict is already on the page.

Data: public-benchmark-shaped slice gaps + seeded synthetic drift. Not Descript's numbers.

3 of 3 quality metrics drifted past their bound last quarter. Under metric-SEV, 3 would have paged within 24h. Under a dashboard watching the average, all 3 waited for the retro.

3
Quality metrics that drifted past their bound
3
Would have paged within 24h under metric-SEV
3
Waited for the retro under a dashboard
Scenario

Scenario 1

Accented-slice WER drifts while the average stays flat

Week 0 / 13
Inside bound. No page.

Dashboard baseline vs metric-SEV, this scenario

Mean metric on the dashboardflat, looks fine
Worst slice / true signaldrifting
Caught under a dashboardnext quarterly retro
Caught under metric-SEVpaged in 24h

On-call view

What fires, and what the runbook says

Hit "Run the worst-case quarter" to watch the metric drift toward its bound. If it breaches, the SEV card and runbook open here.
● SEV-1PAGES ON-CALL
Metric
Bound
Observed
Fired
Blast radius
Response
Runbook — open
Generated post-mortem template