Assurance
Realism is not fidelity.
Synthetic data earns its place in model development only when there is evidence that it represents what the sensor would record — and a clear statement of where it does not. This is how we produce that evidence.
Assurance
Synthetic data earns its place in model development only when there is evidence that it represents what the sensor would record — and a clear statement of where it does not. This is how we produce that evidence.
01 — Our position
A synthetic image or recording can convince a person and still differ from real sensor data in ways a model will learn. A model can even learn to tell real data from synthetic data instead of learning the task.
We therefore treat every dataset as a claim that needs evidence: what was modelled, how it was checked, and where it should not be relied upon.
02 — Four commitments
01
The models, parameters and simplifying assumptions behind each dataset are recorded, so a reviewer can see what the data can and cannot represent.
You receive: a physics summary and parameter record.
02
Where measured data is available, we quantify how far synthetic data sits from it. Where it is not, we say so plainly.
You receive: a validation report with the metrics behind it.
03
Provenance, intended use, out-of-scope use, coverage and known limits, in a consistent structure your assurance team can cite.
You receive: a data card for every dataset version.
04
Held-out synthetic test sets, organised by condition, for probing robustness and edge cases before field trials.
You receive: test sets and a coverage report.
03 — Measuring the gap
No single metric is enough. We combine statistical and model-based measures, and report each one per condition rather than as a single headline figure.
How models trained on synthetic data perform on held-out measured data.
How far synthetic and measured data sit apart in a learned feature space.
How easily a classifier can tell them apart — a warning sign of artefacts a model could exploit.
Which conditions, targets and sensor settings a dataset covers — and which it does not.
04 — The data card
Each dataset version ships with a data card in a fixed structure, so it can be compared, audited and cited in an assurance case.
The example opposite is illustrative; real cards also link to the full physics record and validation report.
LE-EOIR-0003 · Small craft, coastal, LWIR
Version 1.2 · Issued [date]
05 — Your obligations
Synthetic test sets and coverage reports that slot into your test and evaluation planning.
Documentation designed to support assessments under Defence's responsible AI guidance and comparable frameworks.
Data cards that feed directly into the model and system documentation your program maintains.