Evidence validity
A changed model can make part of an existing evidence record stale.
Fionn Labs
Assurance for AI that changes
Applied research for aerospace and defense programs working through a difficult question: what happens to the assurance case when a learned model changes?
NASA X-59 research aircraft. Image: NASA
Applied AI assurance research
Aerospace and defense
Washington State, United States
The assurance gap
The model changed. The assurance case has to catch up
Traditional assurance practice is built around controlled software baselines. Learned components create a harder lifecycle question when they are retrained, compressed, or replaced.
Configuration control
Reviewers need to know exactly which model, property, and conditions a result covers.
Unresolved cases
Timeouts, unsupported cases, and incomplete runs must remain visible rather than becoming implied passes.
Decision boundaries
Component research cannot stand in for a whole-system safety case or an authority decision.
What exists today
What the current work supports
Fionn Labs is building a reproducible research environment around public benchmark materials. Configurations, results, and protected implementation details are not published here.
- Public inputs
- Non-proprietary materials provide a reviewable research context
- Bounded claims
- Conclusions stay within the evidence actually examined
- Private method
- Research architecture and implementation remain confidential
Research discipline
Conservative by design
The public principle is simple: uncertainty is a result to manage, not a gap to hide. The underlying technical implementation remains confidential.
Formal evidence decides
Statistical signals can guide research, but they do not clear a safety claim.
Unknown stays unknown
A timeout, unsupported case, or inconclusive result is withheld rather than softened into a pass.
Change is controlled
Every model version is treated as a configuration event with its own reviewable record.
Claims stay bounded
Evidence is reported only for the property, model, and conditions that were actually examined.
Current maturity
Research-stage, with clear boundaries
What exists today
- Offline, predeployment research using public materials
- Source-grounded analysis of aviation AI developments
- A conservative policy for inconclusive results
- Research and independent replication work in progress
What is not being claimed
- Certification or regulatory approval
- A full-aircraft safety case
- A production deployment decision
- A statistical substitute for formal evidence
Field context
Grounded in the public record
Fionn Labs follows official aviation AI work without presenting a draft, proposal, or research direction as settled policy.
FAARoadmap for Artificial Intelligence Safety AssuranceThe FAA's current framing for AI safety assurance research and certification readiness.EASAArtificial Intelligence Concept Paper, Proposed Issue 3A proposed 2026 update extending the agency's AI assurance concepts; consultation remains open.EUROCAEWG-114 Artificial IntelligenceThe official working-group page lists ED-324 as a draft deliverable.SAEToward a Machine Learning Development LifecycleSAE's public technical paper identifies the companion work as AS6983.
Source review: July 12, 2026
Founder-led
Research informed by program experience
Reid McGill
Fionn Labs is an independent U.S. research company based in Washington State. Its perspective combines commercial aviation digital engineering, doctoral AI assurance research at George Washington University, and U.S. Army Signal experience.
About the companyResearch and program conversations
Start with the program context
We can begin with the program context, the decision you need to support, and the boundary of what can be discussed safely.