Definition
Dependency Discovery Coverage Rate (DDCR) measures the share of actual service dependencies that were active during the AI discovery observation window.
The Measurement Gap
A platform engineering team has deployed an AI-driven service discovery platform across their environment. The platform observes service-to-service communication over a 7-day window and produces a dependency map. The map looks comprehensive: hundreds of services, thousands of edges, automatic categorization by service type. The team uses the map to drive change risk assessment, DR planning, and chaos targeting. The platform reports map completeness above 90%.
Then a routine production change to a financial reporting service causes an unexpected incident in a regulatory compliance reporting pipeline. The two services are not connected in the dependency map. Post-incident investigation reveals that the regulatory pipeline depends on the financial reporting service through a monthly batch job that pulls data on the 5th of each month. The dependency was not in the 7-day discovery window because the change occurred on day 12 of the month, when the batch job had not run since day 5 and would not run again until the next month.
The dependency is real, operationally important, and entirely invisible to the AI discovery because the observation window was shorter than the dependency's activation period. The team investigates further and finds similar gaps: dependencies that fire during DR testing only, dependencies that activate during annual compliance reporting, dependencies that exist for emergency-only paths that don't fire under normal operation. The discovery platform's "90% completeness" is over the dependencies that fire during its observation window, not over the actual dependencies that exist.
The legacy dependency mapping metrics report a healthy state.
Services in catalog: 100% of known services.
Dependencies discovered: thousands.
Auto-discovery completeness: 92%.
Discovery freshness: updated continuously.
Dependency confidence: tracked, mostly high.
None of these report on whether the actual dependency surface is being captured. The completeness measure is observed-vs-observable, not observed-vs-actual. The actual dependency surface is unknown because nobody has enumerated it independently.
These metrics were designed when dependency maps were curated manually. The architect who built the map remembered the monthly batch dependency because they knew the system. With AI discovery, the institutional memory is no longer in the loop, and dependencies that don't fire during observation are not discovered.
There is no metric that tells you what percentage of actual dependencies were active during the AI discovery observation window, which means the map looks complete but is missing every path that didn't fire during observation.
Formula
Dependency Discovery Coverage Rate (DDCR) measures the share of actual service dependencies that were active during the AI discovery observation window. In plain language: independently compile a comprehensive list of dependencies for a sample of services. Compare that list against what the AI discovery captured. DDCR is the percentage of comprehensive dependencies that the AI captured. A high value means the discovery is catching most of what exists. A low value means significant dependencies are invisible.
The formula:
DDCR = Dependencies_discovered / Dependencies_comprehensive × 100
Where:
- Dependencies_discovered = the count of dependencies identified by AI auto-discovery
-
Dependencies_comprehensive = the count of dependencies in a comprehensive inventory compiled through architecture review, batch job analysis, DR runbook review, and pipeline configuration auditUnits are a percentage from 0 to 100. 100% means the discovery captured everything in the comprehensive inventory. A worked example. The platform engineering team conducts a comprehensive dependency inventory for 20 Tier-1 services. The inventory includes:
-
Dependencies active in trailing 7 days: 287 (the discovery window)
- Dependencies active monthly only: 34
- Dependencies active quarterly only: 11
- DR-only dependencies (active only during failover): 18
- Annual compliance dependencies: 6
- Emergency-path dependencies (rarely active): 9
Total comprehensive dependencies: 365.The AI discovery captured 271 of these (some of the 287 active-in-7-days were still missed due to observation gaps).
DDCR = 271 / 365 = 74.2%
Per the catalog guidance (above 80% captures most, below 60% unreliable for change risk) this DDCR of 74.2% is in the middle band. The action trigger is at 70%; the team is just above it. The recommendation is to supplement auto-discovery with periodic comprehensive review (annually for Tier 1) and extend the discovery observation window to 30 days to catch monthly-cadence dependencies. Three edge cases need explicit treatment. First, the comprehensive inventory may itself be incomplete; treat DDCR as a lower bound for the true accuracy. Second, weight dormant dependencies higher in the inventory because their impact when missed is often disproportionate (a missed DR-only dependency might be catastrophic during failover). Third, compile comprehensive inventories for Tier-1 services annually; less frequent inventories produce DDCR values that decay between reviews. Cross-references. Downstream of DDCR is essentially most of the AI-quality metrics that depend on the dependency map: Blast Radius Prediction Accuracy (BRPA, GMI-IT-065), Chaos Coverage Gap (CCG, GMI-IT-047), DR Test Validity Index (DTVI, GMI-IT-054). Low DDCR upstream propagates into all of these. The misinterpretation to watch for: a team reports DDCR = 90% (above the threshold) and concludes their dependency map is reliable. They miss that the comprehensive inventory was compiled by reviewing the same documentation the AI discovery had access to, so the inventory and the discovery share the same blind spots. Compile the comprehensive inventory through independent channels (interview architects, walk runbooks, audit pipeline configurations, examine compliance documentation) not by querying the same systems the AI uses.
Proposed formula, pending validation pilots.