Tactical Radio Channel Simulator (TRCS)
Systematic benchmark suite characterizing speech foundation model hallucination, narrow-band telephony vocoder distortion, and tactical entity extraction under defense, maritime, and public-safety channels.
Narrowband Telephony Codecs, Squelch Physics & Hallucination Suppression
Open BenchmarkAutoregressive sequence-to-sequence speech foundation models exhibit severe failure modes when exposed to operational RF channels: non-speech audio triggers 24 to 160 WPM of continuous ungrounded hallucinations, while 8 kHz low-pass filtering and low-bitrate CELP/RPE vocoders cause catastrophic phonetic collapse. TRCS establishes controlled, reproducible evaluations across clean, synthetic noise, tactical squelch, and narrowband telephone codecs.
| Channel Parameter | Standard Specification | Implementation Parameters | Acoustic Physical Phenomenon |
|---|---|---|---|
| Bandwidth Limiting | ITU-T G.712 bandpass | 300 Hz – 3400 Hz (4th-order Butterworth) | Fricative /s/ sibilance collapse (/s/ → /f, h, t/ causing "Sector" → "Hector") |
| G.711 μ-law / A-law | ITU-T G.711 (PCMU/PCMA) | 8 kHz, 8-bit log PCM (64 kbps) | PSTN, marine VHF DSC gateways, base quantization noise |
| AMR-NB | 3GPP TS 26.071 (MR122) | 8 kHz, ACELP vocoder (12.2 kbps) | All-pole linear prediction smearing of nasal vowel anti-formants |
| GSM 06.10 Full-Rate | ETSI GSM 06.10 FR | 8 kHz, RPE-LTP vocoder (13.0 kbps) | Regular pulse excitation distortion on plosive bursts |
| Opus Narrowband | IETF RFC 6716 | 8 kHz, SILK mode (12.0 kbps) | Tactical mesh IP radios, linear predictive noise shaping |
| RF Squelch Tail | FM Discriminator Unmuting | 50–150 ms bandpass noise burst | Transient insertion trigger causing autoregressive hallucination loops |
Verifier Error Covariance & Optimal Stopping (VEC-SCR)
Formal mathematical framework for sequential candidate selection, positive error dependence, and calibrated abstention in budgeted test-time compute.
Law of Total Error Covariance and Monotonic Sequential Stopping Bounds
Theory & ProofsTest-time compute scaling (Best-of-N, tree search, self-correction) relies on automated verifiers and process reward models. Standard majority voting assumes verifier errors are conditionally independent. Under real-world models and prompts, positive verifier error covariance causes catastrophic voting breakdown. We formalize the covariance matrix Σε and prove that calibrated abstention eliminates false acceptance cascades.
| Theoretical Formulation | Mathematical Expression | Operational Consequence |
|---|---|---|
| Theorem 1: Total Error Covariance | σAB = (1-π)σFA + πσFR + π(1-π)(αA-βA)(αB-βB) | Decomposes joint error into false acceptance, false rejection, and marginal disparity components. |
| Lemma 1: Coupling Saturation | κFA = σFA / √(αA(1-αA)αB(1-αB)) = 89.53% | Empirically proves error dependence saturates 89.5% to 93.2% of the maximal Fréchet-Hoeffding upper bound. |
| Theorem 2: Log-Odds Degradation | Δ I = ln(1 + σFA / (αA αB)) = 2.675 nats | Distorts true 59.37% confidence to an inflated, false 95.49% subjective probability under naive voting. |
| Theorem 3: Stopping Monotonicity | τ1* > τ2* > … > τK* | Under asymmetric loss (CFA ≫ CFR), Bellman dynamic programming thresholds decrease monotonically across stages. |
Sovereign APU Bandwidth & Energy Telemetry
Physical profiling of unified memory architectures (122.7 GB unified LPDDR5X-8000 at 256-bit bus width), memory bus saturation, Joules-per-token thermodynamic asymmetry, and full-node datacenter TCO accounting.
256.0 GB/s Bus Saturation and Joules-per-Token Telemetry
Hardware PhysicsAs open-weight reasoning models scale to 70B–120B parameters, datacenter API telemetry and operational expenditure pose severe constraints. We profile sovereign inference on unified APU hardware running Vulkan compute, sampling silicon package power (PPT) and AC wall-socket draw with calibrated high-frequency smart telemetry at strictly $0.00 cloud spend.
| Architecture Dimension | Sovereign AMD APU (Strix Halo) | Datacenter Cluster (8× H100 SXM5 Node) | Efficiency Multiplier |
|---|---|---|---|
| Unified RAM Capacity | 122.7 GB Unified LPDDR5X | 640 GB HBM3 (80 GB / GPU) | Zero PCIe host-to-device bus transfer bottleneck |
| Memory Bus Width | 256-bit (LPDDR5X-8000) | 5,120-bit per GPU | Sustains 84.1%–87.4% bus saturation across model weights |
| Full Node System Power | 108.4 W active wall draw (17.0 W idle) | 6,300 W IT / 8,378 W Grid (Facility PUE 1.33×) | APU draws 77.3× lower peak system power |
| Energy per Token (Decode) | 2.82 Joules / token | 9.86 – 12.32 Joules / token (slot share) | APU is 3.5× to 4.4× more energy-efficient at batch-1 |
| Marginal Electricity Cost | $0.110 per million tokens | $0.384 – $0.480 per million tokens | Sovereign APU operational electricity is 3.5× cheaper |
| External Cloud Spend | Strictly $0.00 / hr | $16.00 – $24.00 / hr spot rental | 100% data sovereignty; zero remote telemetry leaks |
Canonical Publications & Citations
All papers include verifiable mathematical formulations, reproducible artifacts, SHA-256 manifests, and DataCite DOIs attributed to ORCID 0009-0005-7810-5077.
Tactical Radio Channel Simulator (TRCS): Characterizing Non-Speech Audio Hallucination in Speech Foundation Models
Caceres, Cisco (2026). Published by RoamingPigs Lab & Cisco Caceres AI Systems Lab. Evaluates 6 models across 30 narrowband conditions with 1,050 audio fixtures.
Verifier Reliability Under Budgeted Code Generation: Error Covariance and Information-Theoretic Stopping Bounds
Caceres, Cisco (2026). Published in Cisco Caceres Research Working Papers. Proves Theorem 1 (Law of Total Error Covariance) and Neyman-Pearson sequential stopping bounds.
Hardware Systems Physics and Empirical Energy Telemetry of Sovereign APU Inference
Caceres, Cisco (2026). Sovereign AI Systems Research Reports. Profiles 256.0 GB/s bus saturation and Joules/token scaling across 9B to 122B parameter weights.
The Six Publication Readiness Gates (PRG)
Before any benchmark, dataset, research report, or evaluation ledger is published from RoamingPigs Lab, it must pass 100% of the Publication Readiness Gate checks without exceptions or flakiness.
Institutional Affiliation & Research Contact
RoamingPigs Lab welcomes collaboration on verifier theory, speech foundation model evaluation, and sovereign inference compute.
Direct institutional inquiries, preprint correspondence, and peer replication queries: