Skip to main content

​
RadCheck Score Explained

Your RadCheck score is a composite of four components, each contributing up to 25 points for a maximum of 100.

​
Score Components

​
Stall Risk (0-25 points)

Measures patterns that indicate potential hangs or infinite loops. What we check:
  • Recursive call patterns without exit conditions
  • Unbounded retry loops
  • Blocking operations without timeouts
  • Deadlock-prone resource access patterns
Scoring:
  • 25: No stall patterns detected
  • 15-24: Minor concerns (e.g., missing timeouts on non-critical paths)
  • 5-14: Moderate risk (e.g., retry loops without backoff)
  • 0-4: High risk (e.g., detected infinite loop patterns)

​
Silence Gaps (0-25 points)

Identifies unexpected periods of inactivity that may indicate silent failures. What we check:
  • Time between context updates
  • Heartbeat regularity (if configured)
  • Response latency patterns
  • Activity gaps that exceed expected bounds
Scoring:
  • 25: Activity patterns consistent and within bounds
  • 15-24: Occasional gaps, likely benign
  • 5-14: Frequent or long gaps suggesting reliability concerns
  • 0-4: Extended silence periods indicating likely failures

​
Compaction Pressure (0-25 points)

Evaluates context window utilization and memory management. What we check:
  • Context window fill percentage
  • Compaction frequency and efficiency
  • Memory growth patterns
  • Token budget management
Scoring:
  • 25: Context well-managed, compaction healthy
  • 15-24: Approaching limits but stable
  • 5-14: Frequent compaction, potential information loss
  • 0-4: Context overflow risk, aggressive compaction

​
Hygiene (0-25 points)

Assesses configuration best practices and common misconfigurations. What we check:
  • Provider fallback configuration
  • Error handling completeness
  • Logging and observability setup
  • Security configuration (API key management, etc.)
Scoring:
  • 25: Best practices followed
  • 15-24: Minor gaps in configuration
  • 5-14: Several misconfigurations present
  • 0-4: Critical configuration issues

​
Interpreting Your Score

​
Overall Score Bands

RangeBandAction
80-100🟢 GREENMaintain current practices. Schedule periodic re-scans.
60-79🟡 YELLOWReview specific warnings. Address within 1-2 weeks.
40-59🟠 ORANGEPrioritize fixes. Consider adding Sentinel for protection.
0-39🔴 REDImmediate action required. Do not deploy without addressing.

​
Component-Level Analysis

A low overall score with one component dragging it down is different from evenly distributed issues:
  • Single low component: Focused remediation. Fix the specific area.
  • Multiple low components: Systemic issues. Review architecture.

​
Common Findings

​
”Silence gap detected: Xs between context updates”

Your agent went X seconds without updating context. This might be:
  • Normal (long-running computation)
  • Concerning (waiting on a hung external call)
Fix: Add heartbeat logging or reduce timeout thresholds.

​
”Missing retry configuration in provider fallback”

Your provider fallback doesn’t include retry logic, meaning transient failures cascade. Fix: Add exponential backoff retry to your provider configuration.

​
”Context utilization at X%”

You’re using X% of available context window. Above 80% starts impacting quality. Fix: Review context compaction strategy or reduce context size. Run RadCheck regularly and track scores over time: 
# Save to file with timestamp
acme radcheck --format json > "radcheck-$(date +%Y%m%d).json"
A declining score trend indicates reliability degradation before incidents occur.

​
Next Steps

Sentinel

RadCheck shows risk. Sentinel protects against it continuously.