Research: operator conditioned WiFi representation and observability gates from MotionQ

Problem

RuView increasingly targets shared RF representations across rooms, hardware, link sets, and sensor geometry. A weak premise in many domain generalization pipelines is that stronger invariance across heterogeneous observation geometries must improve transfer. MotionQ, submitted 2026-09-10, argues the opposite when geometry changes what task information is physically observable.

Paper: https://arxiv.org/abs/2609.11818

Why it matters

A WiFi link observes a geometry dependent projection of motion. If an unseen deployment removes a task relevant observable component, feature alignment cannot recreate missing information. Aggressively forcing one invariant latent across layouts can therefore erase useful cues.

MotionQ reports 89.2% mean accuracy across six extrapolation tasks and gains of 20.1, 11.0, and 16.9 percentage points over WiGRUNT, UniFi, and GesFi respectively. Under changing available links it reports 12.0 to 18.4 percentage point gains over the strongest WiFi specific baseline. These are paper results, not RuView results.

Current architecture

RuView already has geometry aware sensing work, cross environment evaluation, hardware normalization, RF embeddings, and an open observability issue (#1854). Existing shared encoder plans still risk treating geometry as a nuisance variable rather than part of the observation operator.

Proposed improvement

Add an experimental operator conditioned representation path rather than replacing the existing encoder.

1. Define an ObservationOperator descriptor containing Tx and Rx poses, carrier or band, available link mask, antenna geometry, calibration state, and timing quality.
2. Condition the RF encoder or task head on this descriptor.
3. Compute an observability summary for the requested task from geometry and current link availability.
4. Train with link retention interventions so single link and reduced link observations remain task sufficient when physically possible.
5. Compare against strict invariant, geometry conditioned, and current main baselines.
6. Expose observability loss to the capability envelope and fusion gate rather than silently preserving confidence.

Implementation phases

Phase 1: benchmark only

Reproduce MotionQ on Widar3.0 or an equivalent public split using the paper protocol. Record exact dataset revision, preprocessing, three or more random seeds, GPU, runtime, memory, and variance.

Phase 2: RuView representation ablation

Use the existing RuView encoder and add only operator conditioning plus link retention. No new backbone dependency.

Phase 3: downstream transfer

Evaluate presence, motion or activity on held out rooms, layouts, devices, link subsets, and people. Do not promote a gesture only improvement into the shared encoder.

Expected measurable improvement

Promotion requires at least one of:

• at least 10 percentage point reduction in cross layout accuracy loss versus current main
• at least 25% reduction in catastrophic high confidence errors when links disappear or geometry changes
• equivalent accuracy with at least 2x fewer target adaptation samples

Nominal known layout performance may degrade by no more than 3 percentage points.

Dependencies

Prefer existing RuView geometry, RF frame, encoder, and benchmark code. RuVector may store operator descriptors and benchmark receipts. No new runtime model dependency is justified initially.

Security and privacy

• Tx and Rx geometry can disclose building layout and location, so treat precise operator metadata as sensitive spatial data.
• Reject malformed and nonfinite coordinates, impossible transforms, oversized link masks, stale calibration, and untrusted frame identifiers.
• Operator metadata cannot increase upstream measurement confidence.
• Prevent dataset metadata or paper text from entering agent instructions without sanitization.
• No network export of raw P0 RF data is required for the experiment.

Compatibility

Experimental feature flag only. Existing encoder and wire outputs remain unchanged. Operator metadata should be additive and versioned.

Testing

• unit tests for geometry and link mask validation
• property tests for permutation stable link descriptors
• regression tests showing current path is byte and metric stable when the feature is off
• adversarial tests for missing links, mirrored geometry, stale calibration, NaN and infinity, timestamp disorder, and mismatched coordinate frames
• x86 and ARM64 inference checks if the conditioning path is promoted

Benchmark plan

Compare four arms with identical data budgets:

1. current main
2. strict invariant alignment baseline
3. geometry conditioning only
4. geometry conditioning plus link retention and observability gating

Measure accuracy, F1, ECE, false high confidence rate, adaptation samples, p50 and p95 latency, memory, model size, and cross environment degradation. Use at least three seeds and report variance.

Rollout

Research only until reproduced. If the RuView ablation passes, promote behind operator_conditioning and observability_gate feature flags. Keep current main as rollback.

Definition of done

• paper result independently reproduced or discrepancy documented
• RuView four arm benchmark complete
• held out geometry and missing link tests complete
• security review complete
• ADR created before any shared encoder contract changes
• no production claim until real multi room and multi device evidence passes

Production classification

Experimental. The underlying paper is fresh and gesture focused; the architectural result is promising but must survive RuView tasks and real deployment splits.

Locate the existing geometry, RF frame, encoder, and benchmark entry points, then start with the Phase 1 MotionQ reproduction on Widar3.0 or an equivalent public split. Compare the four specified benchmark arms with at least three seeds and record the requested metrics, while using the testing section to define validation coverage. Done requires the reproduction or discrepancy report, RuView benchmark, held-out geometry and missing-link tests, security review, and an ADR before shared encoder changes.