[PIR][WP34] Electromagnetic world model via privileged-modality distillation (ADR-338, stretch, ADR-only)

Part of the PIR program epic #837. See 12-wave5-evidence-review.md and 13-wave5-program-plan.md. ADRs land via #911.

ADR-only this wave — ADR-332's deferral is honoured, not re-litigated

ADR-332 (merged, Wave 4) says verbatim: "Deferred implementation. No code lands in this wave." This WP records a design thesis so whichever wave implements has something specific to build against.

Honest greenfield assessment

• crates/ruvector-mmwave/ is a single 419-line lib.rs — a no_std UART frame parser for the Seeed MR60BHA2 protocol, surfacing Event::{Breathing, HeartRate, Distance, Presence, Unknown}. It is a byte-level protocol decoder, not a sensing model. No CSI, no learning, no embedding.
• The only actual CSI code is one bridge binary: crates/ruvector-hailo-cluster/src/bin/ruview-csi-bridge.rs.
• Distillation infra exists but is sensor-unwired — crates/ruvllm/src/qat/{distillation,reasoning_loss,training_loop}.rs is real teacher→student code for LLM quantization. There is no RGB/LiDAR ingest anywhere, no privileged-modality training loop, no CSI dataset loader.
• crates/ruvector-agent-memory/src/observation.rs already enumerates SourceKind::RuViewRf — the provenance vocabulary to reuse.
• ruvnet/RuView is the primary implementation home. It is not checked out locally, despite ruvector-mmwave's docstring citing a ~/projects/RuView/firmware/ path.

The thesis: don't reproduce the paper literally

EMWM (arXiv:2608.17769, verified grade A — SGCS 0.9699, zero-shot at 28 GHz, both exact) requires multi-view RGB at inference, which defeats a substantial part of the RuView deployment thesis.

Instead: RGB / LiDAR / depth as training-only privileged evidence.

CSI + LiDAR + RGB (training)
        ↓
 EM world representation
        ↓
      teacher
        ↓
   distillation
        ↓
  CSI-only student  →  geometry + motion + channel

Intended result, stated so it can be falsified: a camera-free electromagnetic world model trained using temporary visual supervision. A deliberate divergence from the paper, evaluated as such.

Sizing — the dominant cost is data, not code

No code AND no dataset. The arXiv Comments field is empty; GitHub searches return nothing attributable to the authors. The paper's dataset "is constructed based on a campus digital twin" — the authors' own construction, unreleased. Reproduction requires rebuilding a digital twin from scratch. Sizing this as an integration effort would be wrong by an order of magnitude.

Gates for whenever this is implemented

• Re-verify before implementing. ADR-332 already requires re-verification of its source (arXiv:2608.20322, under review at IEEE Access, not peer-reviewed). Same applies here — EMWM's numbers are unreproducible today for want of code and data.
• Training-only provenance is enforced, not assumed. A deployed inference path that acquires a privileged-modality dependency is a regression; the provenance records make it detectable.
• External grounding (ADR-324): ADR-332's designated open corpus (gitlab.ilabt.imec.be/datasets/Activity-recognition-datasets, verified live) satisfies this by construction — a better starting point than the EMWM paper's unreleased twin.
• Expect the student to underperform 0.9699. That figure was achieved with RGB at inference. The acceptance bar is a CSI-only baseline, not the paper's number.

Carry ADR-332's corrected attributions

IR-UWB 89.0% cross-subject F1 / 78.5% unseen-room / €14; FMCW 83.4% / 83.8% best unseen-room / €20; Wi-Fi 79.0% / 68.8% / €320 (SDR dev hardware, not commodity CSI gear). Wi-Fi's 92.6% sleep score is the floor of a three-way near-tie, not a headline. An earlier briefing had FMCW and IR-UWB swapped.

Citation discipline

Cite as "EMWM (arXiv:2608.17769)" in full — "world model" is heavily overloaded (JEPA, Genie, Dreamer) and "EMWM" is an unregistered acronym. Do not adopt either as a crate or module name. Note the paper is eess.SP only, so it will not surface in ML-venue sweeps.

Repos: ruvnet/ruvector docs only. ruvnet/RuView primary future home — coordination required first.

Start with docs/research/perpetual-intelligence-runtime/12-wave5-evidence-review.md and 13-wave5-program-plan.md, then read ADR-332 and the cited crates, especially crates/ruvector-mmwave/lib.rs, crates/ruvector-hailo-cluster/src/bin/ruview-csi-bridge.rs, and crates/ruvector-agent-memory/src/observation.rs. Coordinate with ruvnet/RuView before documenting the design. Done means an ADR-only record of the training-only privileged-modality thesis, provenance gates, verification requirements, and acceptance expectations, with no implementation or dataset claimed.