research: benchmark environment-level RF world embeddings against MERIDIAN site adaptation

Problem

ADR-027 MERIDIAN treats the environment primarily as a nuisance domain to factor out of pose features. New work suggests the environment should also be modeled explicitly as a reusable latent object.

Channel2World (arXiv:2608.17544, submitted 2026-08-18) learns a fixed environment embedding from multiple MIMO channel and position observations. The frozen environment encoder then conditions localization, sparse beam-domain CSI reconstruction, and RF-observable geometry reconstruction in unseen environments without gradient-based site fine-tuning.

This does not invalidate MERIDIAN. It changes the decomposition:

• human/task features should remain environment-invariant where possible
• propagation/environment state should become an explicit reusable world prior rather than discarded training residue

The paper is simulation-only and no official code was verified, so this should begin as a bounded reproduction and benchmark, not a production dependency.

Current architecture

ADR-027 uses a HardwareNormalizer, domain factorizer, environment encoder used mainly as a training signal, and geometry-conditioned pose inference. It currently proposes resampling heterogeneous hardware into a canonical 56-subcarrier representation.

RuView also now has:

• multistatic CSI time alignment
• active sensing and information-gain packages
• spatial/world-model packages
• RuVector-backed memory paths
• iPhone LiDAR geometry capture for future ground truth and fusion

Observed limitations

1. Environment information is treated mainly as shortcut information to suppress rather than a reusable physical prior.
2. Canonical subcarrier resampling can destroy bandwidth, phase, antenna, and hardware information needed for multipath/path-token learning.
3. Site adaptation is still framed around task labels or environment-specific adaptation rather than conditioning from unlabeled channel observations.
4. There is no explicit benchmark comparing environment embedding against fine-tuning at equal context budgets.

Research evidence

Primary: Channel2World, https://arxiv.org/abs/2608.17544

Reported setup:

• 26,000 ray-traced base-station-centered environments
• approximately 5,000 channel measurements per environment
• 128 context channels and 32 query channels in the reported configuration
• 16 latent environment tokens, embedding dimension 128
• approximately 9.2M model parameters
• downstream evaluation on localization, sparse beam CSI reconstruction, and RF-observable geometry

Important boundary: the paper validates on ray-tracing data. It explicitly leaves measured multi-site validation as future work. No production claim should be derived from the reported results.

Related deployment review: Wireless Physical-Layer Foundation Models, arXiv:2608.20486, emphasizes held-out sites, bands, arrays, hardware, mobility, latency, energy, fallback, and hybrid classical/learned operation.

Proposed architecture

Add an experimental environment-conditioning path beside MERIDIAN rather than replacing it.

native complex RF observations + sensor poses
            |
            v
   multipath/path extractor
      AoA / delay / gain
            |
            v
   EnvironmentEncoder
            |
            v
   RfEnvironmentEmbedding
      frozen reusable prior
            |
      +-----+------------------+
      |                        |
 localization head       CSI reconstruction head
      |                        |
 pose/task head           RF geometry head

Store the environment embedding as residual latent state in RuVector. Authoritative geometry, calibration, provenance, and device state remain symbolic and receipt-backed. The latent must never replace known facts.

Target packages

Prefer existing packages. Confirm exact ownership before code changes.

Likely areas:

• MERIDIAN / domain-generalization training path
• existing CSI encoder package
• ruview-memory or current RuVector integration for environment embeddings
• ruview-groundtruth for measured geometry labels
• ruview-bench or current benchmark harness

Do not create a new repository.

Data contract

Prototype:

pub struct RfEnvironmentEmbedding {
    pub environment_id: String,
    pub vector: Vec<f32>,
    pub context_event_ids: Vec<String>,
    pub context_count: u32,
    pub model_id: String,
    pub calibration_ids: Vec<String>,
    pub created_ns: u64,
    pub uncertainty: f32,
}

The authoritative input remains native complex RF plus explicit geometry. Do not make canonical 56-subcarrier data the only retained representation.

Benchmark design

Compare four adaptation strategies under identical context budgets:

A. current frozen baseline

B. MERIDIAN geometry conditioning

C. site-specific fine-tuning

D. frozen environment embedding conditioning

Context budgets: 16, 32, 64, 128, 256, and 1024 observations where available.

Hold out complete rooms, buildings, hardware families, antenna layouts, days, and people where relevant.

Primary metrics:

• localization p50/p90 error
• CSI reconstruction NMSE
• environment geometry error where ground truth exists
• adaptation wall-clock time
• GPU memory
• model size
• context bytes
• inference p50/p95 latency
• calibration error / selective risk

Real-world validation

Synthetic Sionna/ray-traced data may be used for pretraining only.

The promotion gate must use captured or live real data from at least:

• 5 unseen rooms
• 2 buildings
• 3 RF hardware families if available
• 2 antenna layouts
• multiple capture days

The new iPhone LiDAR path can provide metric geometry for a subset of rooms. It must be used as ground truth or calibration evidence, not silently as a runtime feature when evaluating RF-only performance.

MetaHarness configuration

Use the repository-native optimization manifest. Only mutate objectively benchmarkable parameters such as context count, token count, latent width, masking ratio, and adapter size.

Hard vetoes:

• any cross-environment regression beyond the agreed bound
• non-disjoint train/test environment IDs
• higher privacy exposure
• loss of provenance coverage
• latency or memory outside the deployment budget
• synthetic-only gain presented as field gain

Darwin optimization is allowed only after the deterministic benchmark and data-split manifest are fixed.

Security and privacy

Environment embeddings can encode occupancy patterns, room identity, device fingerprints, and potentially sensitive location information.

Controls:

• derive embeddings at the edge by default
• scope them to tenant and physical site
• never allow a raw RF frame to cross a P0 boundary merely because an embedding is lower class
• bind embeddings to source-event receipts
• model poisoning tests using adversarial/malformed RF windows
• bounds on context count and tensor dimensions
• reject NaN/Inf
• model artifact hashes and signed weights
• no external text from research datasets enters agent prompts as trusted instructions

Dependencies and licensing

Do not copy Channel2World code unless an official repository and compatible license are verified. Initial reproduction should be clean-room from the paper using existing RuView/RuVector components.

Expected measurable benefit

Success is not “better embeddings.” It is one or more of:

• equal or better unseen-room localization with at least 4x fewer labeled adaptation examples
• at least 50% reduction in site adaptation wall-clock time
• better sparse CSI reconstruction at a fixed pilot/context budget
• useful RF-observable geometry prediction on held-out real rooms

No benefit is claimed until measured.

Backward compatibility

Experimental feature flag only. Existing MERIDIAN path remains default. No wire-format change is required for the initial benchmark.

Rollback

Disable the environment-conditioning feature and delete its stored latent artifacts. Existing symbolic geometry, RF frames, and MERIDIAN models remain valid.

Definition of done

A reproducible report compares A/B/C/D on fixed disjoint splits with at least five runs for stochastic training, includes hardware/OS/compiler/runtime/model/dataset/configuration, and demonstrates whether environment conditioning improves a real held-out site metric without violating latency, memory, privacy, or provenance gates.

First confirm ownership of the MERIDIAN training path, CSI encoder, ruview-memory or RuVector integration, ruview-groundtruth, and ruview-bench packages. Read the existing benchmark harness and repository-native optimization manifest, then fix disjoint splits and context budgets before implementing the experimental path. Done means a reproducible A/B/C/D report with five stochastic runs and the stated latency, memory, privacy, provenance, and real-site gates.