Launch review
Evaluation report
Draft preview from the current workspace.
This report is a live draft. Run the evaluation to save a timestamped result and unlock baseline comparison.
Overall readiness
69/100
Needs review
Launch decision
Hold for reviewer sign-off before launch.
Model
gpt-4.1-mini
Release
support-prod canary
Risk
High
Baseline
SUPPORT-EVAL-042
Quality gates
| Dimension | Weight | Review signal | Score | Gate |
|---|---|---|---|---|
| Reliability | 32% | Instruction clarity and risk controls | 70 | Review |
| Coverage | 26% | Scenario and failure-path breadth | 76 | Review |
| Drift control | 20% | Baseline and repeatability signals | 69 | Review |
| Evidence | 22% | Sources, criteria, and auditability | 57 | Block |
Reviewer findings
- 01Evidence is under-specified; include expected outputs, citations, or ground truth criteria.
- 02Drift controls need strengthening with a baseline, threshold, or repeatable comparison rule.
- 03High-risk workflows need stricter refusal, escalation, and validation instructions.
Baseline comparison
Save at least two runs to compare the selected evaluation against its previous local baseline.
Return to workspaceEvaluation prompts
System prompt
You are a support triage assistant. Follow policy, ask clarifying questions when evidence is missing, refuse unsafe instructions, and cite the source of each recommendation.
User scenario
A customer reports that an automated workflow deleted records after a billing plan change. Determine the next action, required evidence, and escalation path.