Definition
PVSL measures the 95th-percentile delay between a physical event being captured by a sensor and the digital twin committing the corresponding state update.
The Measurement Gap
Digital twins are increasingly used for real-time decision support, but the industry measures itself with IT metrics. Uptime, availability, API response time. A twin can return a 200 ms response while serving 90-second-old data; "available" and "current" are different things. Platform uptime and OEE cannot see this, because they describe whether the system is running, not whether it is right right now. There is no standard operational metric that decomposes the four-segment delay (sensor capture, network transport, compute/transform, twin commit) end-to-end. When the twin informs operator actions, maintenance scheduling, or autonomous control loops, the latency between physical reality and virtual representation is the difference between a decision tool and a rearview mirror. The tail latency (not the average) is what determines whether the twin is trustworthy during transients, startups, load changes, and fault events, which are exactly the moments when accuracy matters most.
Decision Supported:
Whether the digital twin is fast enough to support a given real-time decision class (operator dashboards, autonomous control loops, transient response), or whether its latency floor disqualifies it from that use.
Formula
Plain language: For every state update, log the sensor capture time and the twin commit time. Subtract to get per-event latency. Take the 95th percentile across all events in the window.
Notation:
PVSL = P95( t_twin_commit − t_physical_capture )
Variables:
| Variable | Definition | Unit |
|---|---|---|
| t_physical_capture | Sensor-side timestamp when the physical measurement was recorded (edge timestamp preferred; IoT gateway receipt is acceptable proxy) | Timestamp |
| t_twin_commit | Timestamp when the digital twin platform committed the state change to its internal model (not queue arrival) | Timestamp |
| P95 | 95th percentile of the per-event latency distribution across the evaluation window | Seconds |
The total can be decomposed into four segments (ingestion, transport, compute, and commit latency) to localize the architectural bottleneck once an overall PVSL value is established.
Proposed formula, pending validation pilots.