Definition
Proportion of the codebase's actual technical debt surface that AI debt-detection tooling can see, measuring the streetlight effect that concentrates remediation where the tools shine.
The Measurement Gap
Measurement Gap and Decision
The Gap: This is the "streetlight effect". Looking for lost keys under the lamp because that's where the light is. AI tools are the streetlight. The team's debt management gravitates toward what tools can see, leaving the most damaging debt unmanaged. The problem compounds because architectural debt's cost is nonlinear: syntactic debt adds linear friction while architectural debt compounds exponentially with each new feature built on the wrong foundation.
The Decision:
Is the engineering team's debt backlog representative of actual debt, or is it biased toward AI-detectable items while structurally important debt goes unidentified? This is a coverage metric: it measures completeness of your AI tools' debt detection vocabulary, not the volume of debt detected.
What Changes When You Have This Number:
DDCR_w above 0.60 means AI covers the majority of important debt types. Supplement with targeted human reviews. Between 0.40–0.60, the debt backlog is materially biased; schedule quarterly human architecture reviews. Below 0.40, the AI-detected backlog is not representative of actual debt and critical types go unmanaged. The metric converts an invisible problem ("we don't know what we don't know") into a visible action plan for each uncovered category.
Formula
Unweighted Formula:
DDCR = Debt_types_AI_detected / Debt_types_total
Importance-Weighted Formula:
DDCR_w = Σ (w_t x detected_t) / Σ w_t
Where Debt_types_AI_detected is the number of taxonomy categories where the AI tool reliably identifies at least 50% of known instances, Debt_types_total is total categories in your agreed-upon taxonomy, w_t is importance weight per type (blast radius, remediation cost, developer friction), and detected_t is binary (1 if reliably detected, 0 otherwise).
Worked Example:
Using the 10-category reference taxonomy with importance weights (1-3 scale):
| Debt Type | Weight | AI Detects? | Weighted |
|---|---|---|---|
| Structural complexity | 1 | Yes (1) | 1 |
| Code duplication | 1 | Yes (1) | 1 |
| Dead code | 1 | Yes (1) | 1 |
| Wrong abstractions | 3 | No (0) | 0 |
| Violated invariants | 3 | Partial* | 0 |
| Implicit coupling | 3 | No (0) | 0 |
| Missing decision docs | 2 | No (0) | 0 |
| Test gaps | 2 | Partial* | 0 |
| Infrastructure debt | 2 | Partial* | 0 |
| Performance debt | 3 | No (0) | 0 |
*Partial detection (<50% of known instances) scores 0.
DDCR = 3 / 10 = 0.30
DDCR_w = 3 / 21 = 0.143
The weighted score is far worse because the three detected categories (complexity, duplication, dead code) all have low importance weights, while the high-impact undetected categories (wrong abstractions, implicit coupling, performance debt) carry weight 3.
Proposed formula, pending validation pilots.