32 lines
1.3 KiB
Markdown
32 lines
1.3 KiB
Markdown
---
|
||
name: guardrail-scorecard
|
||
description: Framework for defining, monitoring, and enforcing guardrail metrics across
|
||
experiments.
|
||
---
|
||
|
||
# Guardrail Scorecard Skill
|
||
|
||
## When to Use
|
||
- Setting non-negotiable metrics (stability, churn, latency, compliance) before launching tests.
|
||
- Monitoring live experiments to ensure guardrails stay within thresholds.
|
||
- Reporting guardrail status in launch packets and post-test readouts.
|
||
|
||
## Framework
|
||
1. **Metric Inventory** – list guardrail metrics, owners, data sources, refresh cadence.
|
||
2. **Threshold Matrix** – define warning vs critical bands per metric / persona / region.
|
||
3. **Alerting & Escalation** – map notification channels, DRI, and decision timelines.
|
||
4. **Exception Handling** – document when guardrail overrides are acceptable and required approvals.
|
||
5. **Retrospective Loop** – log breaches, mitigations, and rule updates for future tests.
|
||
|
||
## Templates
|
||
- Guardrail register (metric, threshold, owner, alert channel).
|
||
- Live monitoring dashboard layout.
|
||
- Exception memo structure for approvals.
|
||
|
||
## Tips
|
||
- Tie guardrails to downstream systems (billing, support) to catch second-order impacts.
|
||
- Keep thresholds dynamic for seasonality but document logic.
|
||
- Pair with `launch-experiment` to ensure readiness before flipping flags.
|
||
|
||
---
|