Research State-of-the-Art & Benchmarks
Relevant source files
The following files were used as context for generating this wiki page:
Research State-of-the-Art & Benchmarks defines the rigorous validation framework for RuView's camera-free RF perception capabilities. This system establishes baseline performance metrics, validates physics-derived theories, and ensures the accuracy of human pose estimation from WiFi Channel State Information (CSI). Sources: examples/research-sota/README.md, aether-arena/space/app.py
The project categorizes all performance claims using evidence labels to distinguish between theoretical potential and physical measurements. Claims are tagged as MEASURED (reproducible with hardware), CLAIMED (reported results), or SYNTHETIC (simulated physics derivations). Sources: harness/ruview/src/guidance.js:33-35, examples/research-sota/README.md:85-88
SOTA Research Loop
The State-of-the-Art (SOTA) Research Loop consists of pure-NumPy reference implementations that validate core physics and spatial intelligence theories. It covers fundamental RF principles including Time of Arrival (ToA) Cramér-Rao Lower Bound (CRLB), Fresnel zone scattering, and adversarial defense mechanisms. Sources: examples/research-sota/README.md:1-12
Research Domains and Findings
The loop identifies critical performance gains through physics-informed modeling and placement strategies:
| Research Area | Headline Finding | Source |
|---|---|---|
| Placement | 93× sensing-coverage lift from physics-aware antenna placement. | examples/research-sota/README.md:58 |
| Detection | 9.36× intruder-detection lift using pose-PABS closed loop. | examples/research-sota/README.md:59 |
| Anchors | 100% household coverage (1-4 occupants) at N=5 anchors. | examples/research-sota/README.md:60 |
| Signal SNR | ~50% breathing-band SNR cost from body multi-scatterer models. | examples/research-sota/README.md:61 |
| RSSI | RSSI alone preserves 95% of full-CSI person count accuracy. | examples/research-sota/README.md:62 |
| Security | Stoer-Wagner mincut identifies all adversarial spoofs. | examples/research-sota/README.md:63 |
| Quantum | NV-diamond cardiac magnetometry recovers vitals at bedside. | examples/research-sota/README.md:65 |
SOTA Dependency Graph
The following diagram shows the logical flow and dependencies between research folders in the SOTA loop.
This graph illustrates how bedrock physics (01) informs placement (02) and spatial intelligence (03), eventually leading to advanced Bayesian fusion (09). Sources: examples/research-sota/README.md:33-53
AetherArena (AA) Benchmark
AetherArena serves as the official spatial-intelligence benchmark for the project. It provides an open, vendor-neutral platform to compare RF-sensing models using standardized splits and metrics. The system implements a hash-chained, tamper-evident witness ledger to ensure result integrity. Sources: aether-arena/space/app.py:1-12
Benchmarking Protocols
The benchmark evaluates models against standardized datasets and split strategies:
- random_split: Evaluates in-domain performance (includes temporal-adjacency leakage).
- cross_subject: Tests generalization against unseen individuals; treated as the flagship deployment frontier.
- Metric: The primary ranking metric is torso-PCK@20 (Percentage of Correct Keypoints).
- Targets: The system targets 87.2% AP@50 under optimal conditions and 51.8% AP@50 for cross-environment scenarios.
Sources: aether-arena/space/app.py:44-50, plans/phase1-specification/system-requirements.md:63-65
Witness Ledger Verification Flow
This sequence diagram describes the process of submitting and verifying a model within the AetherArena sandbox.
Users submit models rather than predictions, which are then scored in a read-only sandbox. Sources: aether-arena/space/app.py:59-70
Performance & Latency Benchmarking
Performance benchmarking measures the operational efficiency of RuView models, specifically targeting edge deployment on devices like ESP32 and Raspberry Pi. The benchmark-ruvllm.js utility evaluates latency, throughput, and quantization effects. Sources: scripts/benchmark-ruvllm.js:1-20
Inference Metrics
The system measures inference latency across multiple percentiles to ensure real-time responsiveness:
- Target Latency: Sub-100ms end-to-end latency on GPU systems.
- Measured Statistics: Mean, P50, P95, and P99 latency.
- Throughput: Embeddings processed per second.
Sources: scripts/benchmark-ruvllm.js:230-247, plans/phase1-specification/system-requirements.md:72
Quantization Scaling
Quantization benchmarks track memory usage and quality loss across different bit-widths:
| Format | Size (KB) | Compression | RMSE | Quality Loss |
|---|---|---|---|---|
| fp32 | Baseline | 1x | 0.000000 | 0.000% |
| 8-bit | ~25% | 4x | Measured | <0.1% |
| 4-bit | ~12.5% | 8x | Measured | Variable |
| 2-bit | ~6.25% | 16x | Measured | Variable |
Sources: scripts/benchmark-ruvllm.js:275-300
Trust Pipeline & Verification
The Trust Pipeline enforces reproducibility through deterministic proofs and witness bundles. The "Trust Kill Switch" prevents behavioral drift by hashing the output of a reference signal processed through the production pipeline. Sources: plugins/ruview/skills/ruview-verify/SKILL.md:9-15
Verification Procedures
- Deterministic Proof: Feeds a reference signal through the pipeline; any mismatch in the SHA-256 result triggers a failure. Sources: plugins/ruview/skills/ruview-verify/SKILL.md:15-18
- Witness Bundle: Generates an ADR-028/ADR-115 compliant bundle containing test logs, firmware hashes, and a self-verification script (
VERIFY.sh). Sources: plugins/ruview/skills/ruview-verify/SKILL.md:27-37 - PCK Baseline: Pose accuracy must include a mean-pose baseline to account for dataset bias. Sources: harness/ruview/src/guidance.js:88-90
RuView's benchmarking framework ensures that all research findings and performance claims are backed by reproducible evidence and audited against standardized spatial intelligence metrics. Sources: aether-arena/space/app.py:112-115