DSFB-RF Structural Semiotics Engine for RF Signal Monitoring - A Deterministic, Non-Intrusive Observer Layer for Typed Structural Interpretation of IQ Residual Streams in Electronic Warfare, Spectrum Monitoring, and Cognitive Radio

DSFB does not compete with matched-filter banks, CFAR detectors, Kalman filters, or ML-based signal classifiers — it augments them. Those systems continue to operate unchanged. DSFB reads the IQ residual streams they already produce and returns a typed, deterministic, human-readable structural interpretation: whether the signal envelope is drifting outward, whether phase noise is accelerating, whether spectral occupancy is approaching its mask boundary, and whether these trajectories constitute structurally organized precursors to regime transitions. The key insight is that every RF receiver already contains a Luenberger-style observer — in its PLL discriminator, AGC tracking loop, or channel equalizer. These incumbent observers treat the innovation residual as a numerical discrepancy to be minimized: an information-erasure enforced by the Luenberger gain matrix L at every update step. DSFB rejects this assumption. It treats the residual as a high-fidelity carrier of unmodeled structural dynamics, mapping IQ trajectories onto the three-dimensional semiotic manifold (‖r‖, ṙ, r̈) and returning a typed grammar state from its structured topology. DSFB is the observer of the observer: a zero-write-path augmentation that recovers the structural information the incumbent observer systematically discards. This paper instantiates the DSFB Structural Semiotics Engine in the RF domain, with formal mapping of IQ residual objects onto the DSFB grammar, typed reason codes, a provenance-aware heuristics bank, and a policy-governed advisory output layer. The companion Rust crate (dsfb-rf) implements the full pipeline in noₛtd / noₐlloc / zero-unsafe Rust, deployable on bare-metal DSP targets without an allocator, converting scalar threshold-crossings into typed, auditable, human-readable episodes. The empirical demonstration applies the engine to two public datasets under a fixed Stage III read-only protocol: DeepSig RadioML 2018. 01a (synthetic, 24 modulation classes, SNR sweep) and the ORACLE dataset (real USRP B200 captures, 16 emitter instances). Under the fixed operator-facing protocol, the policy-governed DSFB layer reduces raw structural boundary events from 14, 203 to 87 episodes on RadioML (99. 4% compression, 73. 6% episode precision, 95. 1% recall over the 102 ground-truth regime transitions in the Stage III fixed protocol; a separate full-dataset amplitude-domain evaluation reports 4. 7% on a denominator of 528 SNR-bin boundaries, not comparable because the denominators count different objects) and from 6, 841 to 52 episodes on ORACLE (99. 2% compression, 71. 2% episode precision, 93. 4% recall). Episode precision improves 102. 2× on RadioML and 76. 8× on ORACLE relative to raw boundary proxies. The upstream receiver pipeline is not modified. If DSFB is removed, upstream behavior is unchanged. This revision additionally reports (a) a sensitivity study at Wₚred ∈ 3, 5, 7 measured on the full RadioML 2018. 01a GOLD HDF5 file (GOLDXYZOSC. 0001₁024. hdf5, 2, 555, 904 captures), populating the previously-deferred cells of the sensitivity table; (b) a per-class amplitude-domain augmentation across all 24 RadioML modulation classes with a JCGM 100: 2008 Type-A Fisher-information visualisation; and (c) an 80-figure companion bank of schema-preserving contextual residual-trace exhibits spanning eight public RF datasets (RadioML, ORACLE, POWDER, Tampere GNSS, ColO-RAN, ColO-RAN CommMag, DeepBeam, DeepSense-6G), reproducible from a single-file Google Colab notebook on free-tier CPU runtime. Context datasets drive no headline number and are disclosed as schema-consistency exhibits; the Stage III headline claims above are driven only by RadioML and ORACLE.