ADR-281: New modality surfaces — BLE Channel Sounding, delay-Doppler-native tensors, P3162 import, and factorized pose

Field	Value
Status	Accepted — implemented (`adapters.rs` BLE CS + ranging evidence, `tensor.rs::delay_doppler_map`, `frame.rs` P3162 import profile, `heads.rs` factorized pose; 8 new test suites)
Date	2026-07-26
Parent	ADR-273; extends ADR-274
Relates to	ADR-279 (`FieldAxis` native axes), ADR-152 (geometry conditioning intake), ADR-021/263 (radar hardware)

0. PROOF discipline

Grades per ADR-273 §0. Bluetooth SIG cm-level claims vs the ~20–50 cm practical review, OTFS ISAC field trials, IEEE P3162, PerceptAlign's >60 % cross-domain error reduction, and RePos's 10–21 % MPJPE gains are EXTERNAL-UNVERIFIED design inputs. Our numbers below are MEASURED-CODE / MEASURED-SYNTHETIC.

1. BLE Channel Sounding (§2) — likely the fastest path to consumer-scale spatial anchoring

BleCsFrame carries per-frequency-step round-trip tone phases plus optional RTT. Two rules:

Phase-based ranging and RTT are separate evidence sources. ble_cs_range computes both — d_phase = |dθ/df|·c/4π from the unwrapped phase-vs-frequency slope, d_rtt = rtt·c/2 — and cross-validates instead of averaging. Agreement raises confidence; divergence beyond 0.5 m yields RangingAnomaly::Divergent (multipath bias, relay attack, timing fault, or calibration problem) with confidence capped ≤ 0.2. Measured: exact recovery at 1.5/5/12 m (< 1 µm error on clean synthetic phases, 40 steps × 1 MHz); a relay-style RTT inflation to ~51 m against a 5 m phase estimate is flagged, not blended (ble_cs_flags_relay_style_divergence_instead_of_averaging).
The tensor view (BleCsAdapter, nrf54-cs in the registry) never detrends phase — the ranging ramp is the measurement; the preserved ramp is asserted in test.

Single-source evidence (no RTT) is capped at confidence 0.5 — one mechanism alone is never high-trust ranging.

2. Delay-Doppler-native support (§3)

FieldAxis (ADR-279) makes delay/Doppler first-class native axes so OTFS-style captures are stored natively, and RfTensor::delay_doppler_map provides the standard transform for frequency-time tensors: IDFT over bins (→ delay) × DFT over snapshots (→ Doppler). Measured: a synthetic scatterer at (delay 7, Doppler 3) produces a unit peak with < 1e-9 leakage everywhere else. The transform is implemented separably (delay IDFT per snapshot, then Doppler DFT per delay row — O(B²S + S²B) vs the direct form's O(B²S²)), proven equivalent to the direct reference to < 1e-10 and measured 8.3× faster (520 µs vs 4.34 ms at 56×8 in the criterion bench). Rule: derived features may be small, but delay-Doppler maps are not collapsed into scalar motion energy before provenance and local storage.

3. IEEE P3162 synthetic-aperture import (§5)

SyntheticApertureSoundingDataset (frequency range, aperture poses, directional PDP, coordinate system, processing-manifest hash) is the validated import profile — the calibration bridge between measured environments, Sionna-class simulators, and learned RF scene models. Schema + validation only; parsers arrive with the first real dataset.

4. Factorized pose (RePos) + log-age gating

FactorizedPoseHead separates what generalizes from what conditions:

relative skeleton branch reads the environment-invariant content representation (cannot learn room-position shortcuts);
root localization branch reads the geometry-conditioned representation (sensor pose is signal there — the PerceptAlign lesson);
absolute = root + relative (PoseOutput::absolute_joints_m), with calibrated per-joint and root residual σ so every pose output carries uncertainty (ADR-273 item 8).

The leakage experiment (factorized_pose_resists_room_shortcut_leakage): training rooms where room position correlates with body scale (the trap real deployments set), held-out room breaking the correlation — factorized MPJPE 0.0003 m vs monolithic absolute-head 0.2534 m (845× worse), on a toy that isolates the mechanism. MEASURED-CODE for the mechanism; not a pose-accuracy claim.

Budget: the structured pose head is the largest adapter at 740 params vs the 40,856-param backbone (1.8 %) — documented ceiling for structured heads is < 2 % (scalar heads keep the 1 % gate), both asserted in every_head_fits_the_one_percent_budget_at_deployment_config.

Age gating now matches the age-aware-CSI recipe exactly: the freshness gate input is log(1 + sample_age_ms) (encoder::age_feature), giving millisecond and multi-second staleness comparable input scale; the finite-difference gradient check re-proves the backward pass through the changed input.

5. Consequences

Bluetooth/UWB anchors slot in as geometric evidence while WiFi carries ambient activity — the complement strategy, in code.
The Gaussian primitive gained the lifecycle fields the update-loop spec requires (first_seen_ns, doppler_variance, bounded source_receipts lineage merged on fusion) — static structure is distinguishable from transients by lifetime, and every primitive traces to source frames.
Roadmap, explicitly not done: real nRF54 CS capture path, OTFS waveform generation, P3162 file parsing, pose heads on real MM-Fi-style data.

5.3 KiB Raw Blame History Unescape Escape