Native observations reach the prediction interface
The new source connects packet-by-packet native predictions and all 27 calibration derivative columns to the structured uncertainty calculation. The finite qualification passes all 24 generated identities, with exact agreement in native values and derivatives. Nine deterministic component checks also pass, including all three physical models under both processing cases, chunked propagation and the complete interface's required abstention when a nonlinear error bound is missing.
The source accepts a fixed initial state and fitted calibration center, a fixed coordinate frame, training/future query times and original increment inputs. Its API accepts no future receiver positions and no fused navigation. Source receipts bind the input identity, initial state, calibration, kernel/runtime, coordinate frame and query schedule. A mismatched model label or uncertainty configuration is rejected.
Those receipts do not prove that a caller's supplied fit or initial state was estimated using training data alone. A future observed study must verify those dependencies independently. The current checks also do not supply useful global nonlinear prediction bounds or establish the receiver-error assumptions. With those requirements unresolved, the complete interface correctly remains an unsupported scientific prediction rather than a detection.
The finite study charged 216 native calls and 28,800 packet visits, with no optimizer starts. No original recording was read, no archival fit ran, and none of the 53 model preferences changed. The affine contractor now passes all 24 frozen controls. The next generated native test checks its usefulness over the full broad region without changing physical limits.
Implementation: analysis/lll/ilvis0_native_stream_interface.py.
Tests: analysis/tests/test_ilvis0_native_stream_interface.py.