Straight answers
What we do not claim
Infrastructure in regulated environments gets bought after a technical review, not before one. So here is the list we would otherwise make you find. If one of these is the thing you needed, we are the wrong purchase and we would rather you knew now.
✗That regulation requires retaining model state
No framework we have checked against primary text mentions internal state. EU AI Act Article 12 requires automatic recording of events, which means logs. We hold no counsel opinion that retained state is a recognised control, and we will not imply one.
✗That we improve accuracy
Reading the whole document improves accuracy. Our layer is byte-identical by design, so it cannot change an answer in either direction. In our own study the arm running with Revyzor and the arm running without it scored the same, which is exactly what a lossless layer should do.
✗That the compression algorithm is novel
It is not, and it is not the product. A competent engineer reproduces the one-off in a few weeks. What is hard is keeping it correct across three serving engines on their own release cadences, successive GPU generations, and model families with different head counts and positional encodings.
✗That we beat a resident cache
We do not, at any context length, and the gap widens as contexts get longer. If the session is still in GPU memory, your engine's own cache is faster than we are and it is free. We are for after eviction.
✗That the saved state is what an auditor inspects
It is a tensor. The signed credential is the human-readable record. The state is what lets someone resume the session as it was rather than rebuild an approximation of it.
✗That state moves between model architectures
It does not. Head counts, head dimensions and positional encoding differ. What is portable is the layer, not the state: switching model family does not mean re-procuring or re-validating the infrastructure underneath it.
✗That we have production customers
We do not. We are pre-seed. Every number on this site was measured on our own hardware, and where an earlier figure did not survive later measurement we replaced it rather than defending it.
What we do claim
A long-context session that falls out of GPU memory cannot be recreated, only re-approximated. The tools that fed it have returned different data since, the serving stack has moved, and the same chain run twice diverges from itself: on our own hardware, 47 of 600 intermediate steps and 2 of 200 final answers.
There are two routes to an identical session. Pin a fully deterministic serving stack and recompute, which costs roughly half your fleet throughput and still cannot recreate tool outputs that have changed. Or keep the state, at 1.444x lossless on bf16, signed, on your own storage.
That is the whole proposition. Everything else on this site is detail underneath it.
If you think one of the rows above is wrong, we would like to hear it. We have changed published claims on this evidence before.