Research question
When the model changes, what happens to the assurance case?
A model update can make part of an existing evidence record stale. The responsible question is not whether change can be waved through, but how its assurance impact can be understood without overstating the result.
A controlled baseline is more than a file
An assurance case connects claims, assumptions, configurations, and evidence. When a learned component changes, the new weights file is only one part of the change. The update may also affect the behavior being examined, the conditions under which prior evidence was gathered, and the scope of conclusions a reviewer can still rely on.
That is why an accuracy score alone cannot answer the assurance question. It may be useful engineering information, but it does not establish which prior claims remain supported.
What can be said publicly
A model change can affect whether existing evidence still supports the same conclusion. That is the public research problem. The technical process Fionn Labs uses to investigate it is not described on this site.
Sequencing, decision logic, thresholds, campaign design, and evidence structures remain confidential.
What a bounded conclusion sounds like
A credible conclusion states its scope and limitations without extending beyond the evidence reviewed. Component research does not become a whole-system safety claim or substitute for a decision by a certification authority.