Research: packet level complex CIR coherence for communication native mmWave sensing

[ruvnet]

Problem

RuView already preserves WifiCir as an authoritative native RF modality and already has CSI phase sanitation, antenna phase alignment, CIR estimation, multistatic CIR gating, calibration receipts, and sample age. The missing piece is packet to packet temporal coherence for communication native mmWave CIR streams.

A fresh physical testbed result on programmable IEEE 802.11ad exposes the full complex 128 tap CIR for every packet and shows that raw packet timing and common phase drift materially degrade sub wavelength motion sensing.

Source: https://arxiv.org/abs/2609.15622

Across 39 recordings, the paper reports median waveform correlation improving from 0.41 after timestamp regularization to 0.52 after delay alignment and 0.61 after static reference relative phase calibration. Breathing rate mean absolute error improves from 5.1 to 4.1 to 3.6 breaths per minute.

These are measured paper results, not RuView results.

Why it matters

RuView already treats native complex samples, phase state, geometry, calibration, sample age, and provenance as first class through ADR 279. Adding temporal CIR coherence at the correct layer could improve vital sign, micro motion, radar style gating, and future 60 GHz sensing without changing the wire contract or introducing a new model.

Current architecture

Relevant existing components:

1. v2/crates/ruview-unified defines RfFrameV2 and WifiCir.
2. v2/crates/wifi-densepose-signal/src/ruvsense/cir.rs derives CIR from phase sanitized CSI.
3. PhaseSanitizer and phase_align.rs handle existing phase cleanup and antenna alignment.
4. ruvsense/calibration.rs records room calibration with phase preconditions.
5. ruvsense/multistatic.rs already supports CIR based delay gating.

Observed limitation

Current phase sanitation and antenna alignment do not constitute an explicit packet level coherence contract for a native CIR stream. Delay drift, common phase drift, missing packets, reordered packets, stale static references, and reference path disappearance need explicit handling and observability.

Proposed improvement

Add an experimental CirCoherenceCalibrator in wifi-densepose-signal behind a feature flag.

Inputs:

1. Native complex CIR taps from RfFrameV2 or the existing CIR type.
2. Monotonic packet sequence or timestamp.
3. Calibration state and phase state.
4. Optional static reference delay window selected during calibration.

Outputs:

1. Delay aligned CIR.
2. Relative phase calibrated CIR.
3. Estimated delay shift.
4. Estimated common phase drift.
5. Reference path quality.
6. Coherence confidence.
7. Typed abstention reason when calibration is not supportable.

The calibrator must never overwrite the authoritative native CIR. It produces a derived view with a new receipt referencing source frame ids and calibration id.

Implementation phases

Phase 1: deterministic replay

Implement delay profile alignment and static reference relative phase calibration against synthetic and recorded complex CIR fixtures. Add packet loss, jitter, reordering, and static path disappearance perturbations.

Phase 2: RuView downstream benchmark

Compare four arms using the same frozen data:

1. current RuView path
2. timestamp regularization only
3. delay alignment
4. delay alignment plus relative phase calibration

Measure waveform correlation, breathing MAE, micro motion spectral SNR, false confidence, p95 latency, allocation count, and deterministic replay.

Phase 3: physical hardware

Run at least 30 recordings across at least three rooms, two subject distances, two antenna layouts, and multiple packet loss conditions. If programmable 802.11ad hardware is not available, use an equivalent native complex CIR source but do not claim 802.11ad reproduction.

Security and privacy

1. Native complex CIR remains P0 or equivalent raw RF evidence and stays inside the edge trust boundary by default.
2. Reject non finite taps, impossible shape, non monotonic sequence unless explicit reorder buffering is enabled, oversized frames, and unbounded reference windows.
3. Limit memory and CPU for alignment search to prevent resource exhaustion.
4. A stale or disappearing static reference must lower confidence or force abstention, never silently preserve prior calibration.
5. A malicious reflector can perturb the reference path. Record reference path quality and do not treat it as device identity or authentication.
6. Calibration receipts bind configuration, source evidence class, algorithm version, and chosen reference window.
7. No vital sign, safety, occupancy, or medical claim is promoted from replay or synthetic evidence.

Benchmark plan

Baseline against current main on the same recordings.

Record:

1. hardware and firmware
2. OS and Rust compiler
3. centre frequency and bandwidth
4. CIR tap count and packet rate
5. packet loss and jitter
6. waveform correlation
7. breathing MAE where ground truth exists
8. spectral SNR around induced micro motion
9. p50 and p95 processing latency
10. memory and allocations
11. false high confidence rate under reference loss
12. number of runs and variance

Run at least three deterministic replays per fixture and multiple independent physical captures.

Success metrics

Promotion to production candidate requires all of:

1. At least 15 percent relative improvement in waveform correlation over the current RuView path on held out measured captures.
2. At least 15 percent relative reduction in breathing MAE where independent ground truth exists.
3. No more than 2 percent degradation on nominal current CSI or CIR fixtures.
4. p95 incremental latency below 1 millisecond per frame on the reference x86 gateway and below 2 milliseconds on ARM64 edge hardware.
5. Zero false high confidence outputs when the reference path is removed in the adversarial fixture.
6. Byte deterministic replay for the same input and configuration.

MetaHarness plan

Use stable ruvnet/metaharness v0.4.4 with a repository specific evaluator. The objective is multi dimensional: improve measured coherence and downstream error while preserving latency, memory, determinism, privacy, and calibration failure behavior. Darwin optimization is allowed only for bounded parameters such as reference window width or alignment search radius after a frozen holdout exists. Candidate logic may not modify tests or holdout manifests.

Dependencies

No new runtime dependency is required for the initial implementation. Reuse existing complex math, CIR, calibration, frame, provenance, benchmark, and RuVector receipt infrastructure.

Compatibility and migration

Additive feature flagged path only. Native RfFrameV2 and existing CIR representations remain unchanged. Disabling the feature returns current behavior. No stored frame migration.

Rollback

Remove the feature flagged calibrator and derived receipt type. Native RF data and all existing calibration paths remain valid.

Definition of done

1. ADR covering packet level CIR coherence, security, privacy, benchmark design, and rollback.
2. Unit, regression, property, malformed input, and deterministic replay tests.
3. x86 and ARM64 release benchmarks.
4. Real measured capture comparison with independent ground truth where used.
5. MetaHarness receipt bound to the exact tested commit.
6. No claim promoted beyond its evidence level.

Production classification

Experimental until measured multi room reproduction passes the gates above.