Offline and determinism

No network at scan time

Model weights are fetched once, at install or on first load, and cached. After that, scan_input and scan_output work with the network interface down.

This is asserted rather than promised. The test suite patches socket.socket to raise and runs the default suite against it. If a change makes a test need the network, the change is wrong, not the test.

Nothing is sent anywhere. No telemetry, no usage counters, no crash reports. The evidence record stays on your infrastructure.

What that costs

CPU is the reference target rather than a fallback, so every detector has to be usable on a CPU within its budget. GPU is an optimisation, never a requirement. That constrains what the models can be, and it is the reason the classifiers are encoders rather than anything generative.

Determinism

Given the same inputs and the same model revisions, a scan returns the same decision and the same record. No sampling, no temperature, no clock-dependent behaviour inside a scan.

The timestamp in the record is the one thing that changes between runs, and it is deliberately outside the signed comparison you would make when reproducing a decision: hash the inputs and the policy, not the record as a whole.