Definition
Compliance Documentation Currency Index (CDCI) measures the share of compliance-critical model documentation fields that still accurately describe the current model's behavior.
The Measurement Gap
A financial services firm completed a regulatory examination of their AI-driven anti-money-laundering system 18 months ago. The model cards, risk assessments, fairness analyses, and operational documentation were current at examination time. The examiner signed off. The compliance team filed the documentation and moved on to the next regulatory cycle.
In the intervening 18 months, the AML model has been retrained four times. The feature set has expanded twice. The decision thresholds have been tuned three times. The integration with downstream investigator workflows has been changed once. The original model card's description of feature importance is no longer accurate. The risk assessment's discussion of false positive rates is based on the old threshold. The fairness analysis was performed on the original training data; the current training data differs in composition. The operational documentation describes a workflow that has been superseded.
Then the next examination cycle starts. The examiner requests current documentation. The compliance team realizes the gap. They scramble to update the documentation in the two weeks before the examiner arrives. The updates are rushed, partial, and arguably less rigorous than the original. The examiner notices the freshness pattern (recent updates with timestamps suggesting last-minute work) and asks pointed questions about what changed between exams. The conversation goes badly.
The legacy compliance metrics on the firm's dashboard report a healthy documentation program.
Models with model cards: 100%.
Models with risk assessments: 100%.
Documentation completeness: 100% at last audit.
Audit pass rate: 100%.
Documentation freshness at audit: meets standard.
Documentation retention: 7 years.
None of these report on whether the documentation is current at any given moment between audits. The completeness check happens at audit time. Between audits, the model evolves and the documentation does not. The compliance posture decays continuously and is reset to 100% only at audit time, after a rush to update.
These metrics were designed when AI models were stable artifacts that changed rarely. In that world, documentation written at deployment was largely valid for the life of the model. With modern AI systems being retrained frequently, having features added regularly, having thresholds tuned operationally, and being integrated and re-integrated with workflows, the documentation decay rate has accelerated dramatically. The audit-time check has become a snapshot of a continuously-decaying state.
There is no metric that tells you whether compliance documentation still describes the current model's actual behavior, which means audit-time compliance snapshots decay between audits while the model evolves.
Formula
Compliance Documentation Currency Index (CDCI) measures the share of compliance-critical model documentation fields that still accurately describe the current model's behavior. In plain language: take the model's compliance documentation, identify the fields that are operationally critical (the ones a regulator would check), and assess each field against the current state of the model. Count the fields that are still current. CDCI is the percentage that are still accurate. The check is performed quarterly on every regulated model, not only at audit time.
The formula:
CDCI = Fields_current / Fields_total × 100
Where:
- Fields_current = the count of compliance documentation fields that accurately describe the current model
-
Fields_total = the total count of compliance-critical documentation fieldsUnits are a percentage from 0 to 100. 100% means fully current documentation. Lower values mean accumulating documentation debt. A worked example. The financial services firm runs CDCI on their AML model after the rough examination. They identify 38 compliance-critical documentation fields (feature list, feature importance distribution, model type and version, training data description, decision thresholds, performance metrics, fairness analysis results, false-positive rate by segment, operational workflow integration, override authority, audit log structure, retention policy, retraining schedule, monitoring scope, alert thresholds, incident escalation paths, and so on). Assessment of each field against the current model:
-
22 fields current and accurate
- 9 fields stale (the field's content was correct at the last documentation update but no longer reflects the current model)
- 7 fields incorrect (the field's content actively misrepresents the current model)
CDCI = 22 / 38 = 57.9%
Per the catalog guidance (above 90% for regulated models, below 80% indicates material compliance risk) this CDCI of 57.9% is well below the action threshold. The team needs to refresh the documentation across all stale and incorrect fields, and implement an automated check that flags drift between documentation and model state going forward. Three edge cases need explicit treatment. First, define materiality thresholds per field. Some documentation fields tolerate small drift without compliance impact (high-level descriptive text); others require exact accuracy (decision thresholds, regulated disclosures). Mark fields as material or non-material in advance, so the CDCI calculation focuses on what regulators care about. Second, models not retrained since the last audit have trivially-100% CDCI, but the underlying production data may have changed. Track time since last review alongside CDCI so a model with high CDCI but a long review gap gets attention. Third, assess all regulated models quarterly; less frequent assessment misses the drift between assessments. Cross-references. Upstream of CDCI is Remediation Effort Yield (REY, GMI-IT-019). When documentation refresh effort is invested, REY tells you whether the effort actually reduces operational risk. Downstream is Proxy Discrimination Detection Gap (PDDG), a model with high PDDG should have its CDCI flagged because the fairness documentation is incomplete by definition. The misinterpretation to watch for: a team reports CDCI = 95% (above the threshold) and concludes documentation is current. They miss that the 5% of fields out of date are the most operationally consequential ones (decision thresholds, performance metrics, fairness analysis) because those are the fields most likely to change when the model is retrained. CDCI should be reported both overall and weighted by materiality, with materiality-weighted CDCI being the operationally relevant number.
Proposed formula, pending validation pilots.