Definition
Integration Semantic Drift Index (ISDI) measures the cumulative semantic divergence of auto-generated integration mappings from original field-level intent, weighted by integration criticality.
The Measurement Gap
A multinational manufacturer runs an integration platform that connects 47 source systems to 12 downstream consumers (data warehouse, analytics platforms, regulatory reporting, customer portal, ERP). The integration platform has an AI-driven mapping engine that suggests and refines field-level mappings based on observed data patterns. Over the past three years, the engine has remapped fields dozens of times across the integration portfolio, each time validated against immediate-downstream consumer feedback.
Then the finance team raises a question during the quarterly close. The "customer revenue" field in the data warehouse no longer matches the "customer revenue" field in the ERP. Investigation reveals that the integration platform has remapped the source-system field underlying "customer revenue" three times in the past 18 months. Each remap was validated as correct at the time: the downstream consumer accepted the new mapping and used it without complaint. But the cumulative effect of three remaps is that the field now represents "trailing-12-month customer revenue net of returns" instead of the original "trailing-12-month customer revenue gross of returns." Each remap was reasonable in isolation. The cumulative semantic drift was undetected because no single comparison surfaced the cumulative gap.
The finance team has to reconcile two systems that are supposed to mean the same thing and no longer do. The regulatory reporting team realizes their filings have been based on the drifted field for six months. The compliance team starts a documentation review.
The legacy integration metrics report a healthy integration platform.
Integration uptime: 99.97%.
Mapping validation pass rate: 100%. Every mapping change was validated.
Downstream consumer feedback: positive on individual changes.
Schema validation: 100%. Types match, formats valid.
Integration throughput: stable.
None of these report on whether the cumulative semantic intent of an integration mapping still matches the original specification. Each change was validated against the immediate prior version. The mappings drift one increment at a time, each increment within tolerance, until the cumulative drift exceeds any single increment's threshold.
These metrics were designed for static integrations or for integrations whose changes went through human review with a long-term semantic context. The integration architect would notice that a mapping had drifted across three iterations because they had been involved in all three. With AI-driven mapping engines making changes faster than human review, the cumulative semantic drift is no longer in anyone's head.
There is no metric that tells you whether cumulative auto-generated integration remappings have drifted from original semantic intent, which means each mapping passes validation individually while end-to-end data meaning erodes.
Formula
Integration Semantic Drift Index (ISDI) measures the cumulative semantic divergence of auto-generated integration mappings from original field-level intent, weighted by integration criticality.
In plain language: for each auto-remapped integration, compare the current mapping against the original mapping specification. Categorize the comparison as Equivalent (same semantic meaning), Approximate (close but not identical meaning), or Divergent (different meaning). Weight by integration criticality and compute the aggregate drift score.
The formula:
ISDI = Σ (w_i × drift_score_i) / Σ w_i
Where:
- w_i = the criticality weight of integration i (typically 1, 2, or 3 for low/medium/high criticality)
- drift_score_i = 0.0 if the current mapping is semantically equivalent to the original, 0.5 if approximately equivalent, 1.0 if divergent
Units are a weighted ratio from 0 to 1. 0 means no drift across the integration portfolio. 1 means all critical integrations have diverged.
A worked example. The manufacturer runs ISDI on 30 of their auto-remapped integrations. Criticality classification: 8 high (weight 3), 14 medium (weight 2), 8 low (weight 1).
Domain expert review of each integration against original specification:
| Criticality | Equivalent | Approximate | Divergent | Total |
|---|---|---|---|---|
| High (w=3) | 5 | 2 | 1 | 8 |
| Medium (w=2) | 9 | 4 | 1 | 14 |
| Low (w=1) | 6 | 1 | 1 | 8 |
Drift score contribution:
- High: 5×3×0.0 + 2×3×0.5 + 1×3×1.0 = 0 + 3 + 3 = 6
- Medium: 9×2×0.0 + 4×2×0.5 + 1×2×1.0 = 0 + 4 + 2 = 6
- Low: 6×1×0.0 + 1×1×0.5 + 1×1×1.0 = 0 + 0.5 + 1 = 1.5
Total weight: 8×3 + 14×2 + 8×1 = 24 + 28 + 8 = 60
ISDI = (6 + 6 + 1.5) / 60 = 13.5 / 60 = 0.225
Per the catalog guidance (below 0.10 holding meaning, above 0.30 interpretation risk) this ISDI of 0.225 is in the warning band. The action trigger at 0.20 for Tier 1 integrations has been crossed (the High-criticality slice alone has ISDI = 6/24 = 0.25). The team needs to manually re-anchor the High-criticality integrations to original specifications, then add a process for auditing future auto-remapping against the original spec rather than against the most recent prior version.
Three edge cases need explicit treatment. First, original intent must be documented. If no original specification exists, ISDI is not computable (and the absence of the specification is itself a risk indicator). Second, compare to the original specification, not the most recent prior version; drift accumulates across iterations and the comparison must span the full history to detect it. Third, the minimum is 20 auto-remapped integrations per measurement; below that the weighted average is unstable.
Cross-references. Upstream of ISDI is
Schema Propagation Cascade Depth (SPCD): schema changes and mapping changes are related; SPCD measures how far structural changes propagate, ISDI measures how far semantic changes drift. Downstream is
Downstream Impact Propagation Index (DIPI): drifted mappings produce data that reaches consumers in a different meaning than intended, contributing to consumer-side incidents.
Proposed formula, pending validation pilots.