AI Slop Prevention Score¶
The AI Slop Prevention Score is a composite metric that quantifies the quality and trustworthiness of an AI response. It combines 6 weighted components into a single 0.0-1.0 score.
Components¶
| Component | Weight | Range | Description |
|---|---|---|---|
| Grounding Coverage | 0.25 | 0=grounded, 1=ungrounded | Whether output is backed by evidence |
| Schema Compliance | 0.15 | 0=valid, 1=invalid | Whether output matches expected schema |
| Unsupported Claim Ratio | 0.20 | 0-1 | Ratio of claims not found in context |
| Genericity Score | 0.15 | 0=specific, 1=generic | Density of filler phrases |
| Policy Risk Score | 0.15 | 0-1 | Policy violation severity |
| Action Risk Score | 0.10 | 0-1 | Risk level of proposed actions |
Formula¶
slop_score = 0.25 * grounding_coverage
+ 0.15 * schema_compliance
+ 0.20 * unsupported_claim_ratio
+ 0.15 * genericity_score
+ 0.15 * policy_risk_score
+ 0.10 * action_risk_score
Decision Thresholds¶
| Score Range | Decision | Meaning |
|---|---|---|
| 0.0 - 0.3 | pass |
Output is clean |
| 0.3 - 0.7 | repair |
Output needs improvement |
| 0.7 - 1.0 | reject |
Output is unacceptable |
Thresholds are configurable per tenant via SLOP_SCORE_THRESHOLD_PASS and SLOP_SCORE_THRESHOLD_REPAIR.
Source¶
src/agentguard/slop_score/