Sector
Aerospace
- AI testing
- Autonomy
- Digital engineering
Where the certification argument is the deliverable, not the documentation.
Airborne software has the most developed assurance culture of any of these sectors, and it is precisely that maturity which makes learned components difficult: the existing objectives assume a specification you can trace to code, and a learned component does not have one in that sense.
The practical consequence is that the argument carries more weight and the evidence underneath it is scrutinised harder. A dossier that regenerates to a different digest, or a claim that resolves to an artefact nobody can replay, will not survive the first review.
What this means for evidence
Deterministic certification packages
PDF/A-3b with the evidence embedded, provenance-linted, byte-identical across two generations of the same input.
Argument structure with explicit defeaters
GSN argument graphs where every draft node must be disposed of and the sim-to-real gap is a mandatory defeater, not an optional one.
Reproducibility class on every run
R0 to R3 classification with automated diagnosis, and refusal to cite evidence at an inadmissible class.