Phase detection that has to commit
A causal, forward-only decode at one inference per operating-room second. A model that cannot revise its own history is harder to build and is the only kind that is useful in the room.
Most published surgical-phase numbers are retrospective. The model sees the whole case and then labels it. That is a legitimate research setting, and it is not the setting a system in an operating room is in.
A retrospective model can revise its own history. If it decides at minute forty that the case entered surgery at minute twelve, it can go back and say so, and the reported accuracy improves. A system that has to tell a team what is happening now gets no such opportunity. It commits, and it lives with the commitment.
The pipeline
One inference per labelled operating-room second, decoded forward-only, and replayed through a causal stream simulator so the numbers come from the same code path a live install would use.
The forward-only constraint is the whole design. The decode cannot revisit a state it has left. This costs accuracy against a bidirectional decode, and we report the cost rather than quietly evaluating the causal system in the retrospective setting.
Latency
The streaming path is measured at roughly 150 to 250 ms from camera to frame in memory on the theatre network.
That figure is transport and decode, not end-to-end model latency, and we label it that way because the difference is exactly where vendor numbers usually go soft.
Why phase is the anchor readout
Almost every operationally useful question about a case is a question about timing relative to phase. How long between cases. How much traffic crossed the room during the sterile phase. Whether antibiotic timing landed inside the window. Whether the count happened before closure.
None of those are answerable from a frame in isolation. All of them are answerable from a phase track that is honest about when it knew what it knew.
This is the earliest note.
The headline sterility number is a rule, not a model