Paper 5: What survives when the model, client, and machine change?
A model session ending should not mean that the system forgets who is authorized, what is in scope, what evidence supports a record, or how an earlier decision can be audited.
That sounds obvious—until you ask where those things actually live.
If they live only in a live prompt, a model’s context window, a particular client, or the memory of one running process, then they are not really system continuity. They are temporary convenience.
My fifth preprint, The Continuity Substrate, tests a different proposition: can governed operational continuity survive changes in models, clients, machines, operating systems, and runtime restarts when the continuity contract is maintained outside the inference session?
The answer from this controlled study is narrow but encouraging.
Under named first-party synthetic conditions, the governed substrate preserved its deterministic contract across:
- 12/12 frozen fixtures
- 23/23 unchanged kernel tests
- 13/13 security-boundary tests
- encrypted restart recovery, with 3/3 governed records and provenance pointers recovered
- a different read-only client, with zero observed cross-scope leaks
- a 50-client local loopback run: 900/900 staged retrieval operations and 50/50 post-restart recoveries
- a 25-case governance-adversarial suite with zero unauthorized effects
The point is not that any model has become magically permanent. It has not.
The point is that important continuity should not be entrusted to a model session in the first place.
In this architecture, models can change. Clients can change. A runtime can restart. The intelligence can come and go. What must remain stable—authority, scope, provenance, consent, auditability, and the rules for releasing information—belongs in a governed substrate outside inference.
That matters for ordinary software as much as it matters for AI. If someone changes providers, upgrades a model, moves from one device to another, or restarts a service, they should not have to negotiate their digital history from scratch. And they should never discover that a new model has quietly inherited information it was not authorized to see.
There are real limits to this paper. It does not claim production readiness, security certification, universal portability, or independent replication. It reports a controlled test and a public evidence package so readers can inspect the boundary being claimed.
But it asks a question I think the industry needs to take more seriously:
When the intelligence changes, where does the responsibility stay?
Read the open preprint and inspect the public evidence package here:
https://doi.org/10.5281/zenodo.23990079
The model may be brilliant. But continuity, permission, and accountability should not disappear when the model closes its eyes.
