Synthetic Continuity-Residual Fusion Sandbox — JuPedSim + Sionna + BLE Co-Simulation

Motivation

Ch3 L3.4 commits the thesis to the continuity equation $\partial \rho/\partial t + \nabla \cdot (\rho \mathbf{v}) = 0$ as an explicit regularisation loss during CSI inference. That is the architectural novelty of the four-tier system. But:

• Ch5 L5.6 (continuity-as-regularisation-loss reduces variance and improves OOD generalisation) is currently marked missing — 0 papers. The chapter's central claim has no supporting evidence.
• The closest published precedent (di2023_285b ↗, zabin2026_a20c ↗) covers physics-informed losses generally but not the continuity equation specifically against indoor occupancy CSI fusion.

The hardware-driven experiments EXP-001 through EXP-005 all require the FIIT testbed (currently un-mapped, sensors not yet deployed). A simulation-only sandbox lets us:

1. De-risk the architectural commitment. Before investing in a 4-week calibration campaign, confirm that the continuity-residual penalty measurably improves per-cell occupancy MAE under controlled conditions.
2. Isolate the mechanism from the architecture. By running the same penalty across a model zoo (Bagging-LGBM, PINN, CNN+LSTM, HAN-GAT, GCN, TKL), we answer the reviewer-foreseeable question: "is it the physics prior doing work, or is it the deep model?" — empirically, with a 2-D ablation matrix.
3. Build the fusion-training infrastructure now. Graph construction code, model zoo, λ-sweep harness, regime-conditional analysis tooling — all reusable when real data lands.
4. Surface upstream blockers cheap. GPU envelope for Sionna RT, JuPedSim trajectory rate, per-window graph size — discoverable in days, not weeks.

Setup

Simulation Stack

graph LR
    A[JuPedSim] -->|trajectories x,y,t| B[Sionna RT / QuaDRiGa]
    A -->|agent positions| C[BLE path-loss model]
    A -->|trajectories| D[PedPy]
    B -->|CSI per subcarrier per link| E[Feature aggregator]
    C -->|RSSI per beacon-observer pair| E
    D -->|per-cell density, velocity, flow, CV| F[Per-cell occupancy GT + regime label]
    E -->|per-window feature tensor| G[Model zoo]
    F -->|targets + regime| G
    G -->|per-cell N̂ + continuity-residual| H[λ-sweep evaluation]
    F -->|∂ρ/∂t + ∇·(ρv)| H

Synthetic Geometries

Unchanged from the previous EXP-006 design — the geometry layer is independent of the model headline.

Geometry A — single seminar room (~96 m², capacity 12–24):

+----------------------------------------+
|              CELL A (8 × 12)           |
|                                        |
|  [CSIAP-1]               [BLEBeacon-1] |
|                                        |
|                                        |
|  [BLEBeacon-2]           [CSIAP-2]     |
|                                        |
+---[Door-N]----------------[Door-S]-----+

Six target cells (3 × 2 logical grid), 2 CSI AP pairs, 2 BLE beacons, 4 BLE observers at the corners. Mirrors EXP-001's "meeting / seminar room".

Geometry B — corridor with side rooms (~110 m², capacity 20–40):

+----[Room-W]----+                  +----[Room-E]----+
|  CSIAP-W   B-W |                  |  B-E   CSIAP-E |
|                |                  |                |
+------[D-W]-----+    CORRIDOR      +------[D-E]-----+
                  +================+
                  |   (20 × 4)     |
                  |  [CSIAP-mid]   |
                  +================+

10 target cells, 3 CSI AP pairs, 4 BLE beacons, 6 BLE observers. Captures multi-cell topology (room → corridor transitions, door events).

Materials in Mitsuba (Sionna): concrete walls, wooden doors, glass windows. 5 GHz Wi-Fi (channel 36, 80 MHz) for CSI; 2.4 GHz for BLE.

Per-Window Feature Schema

Window: 1 s. CSI 100 Hz (aggregated per window). BLE 1 s advertisement interval (matches EXP-001). Target: per-cell occupancy at $t + 1,\text{s}$.

Phone-Penetration Sweep

Same scenario tree as EXP-005:

with $N \in {4, 8, 16, 24}$ for Geometry A and $N \in {8, 16, 24, 40}$ for Geometry B.

Procedure

Phase 0 — Synthetic geometry + pipeline plumbing

1. Author Geometry A and B as Python factories. Versioned under experiments/exp006/geometry/.
2. Stand up QuaDRiGa-fast-tier channel via csi(geometry, agent_positions, link_pair) → H[subcarrier]. No Sionna RT yet. Goal: flush bugs through the data flow.
3. Run end-to-end on the tiniest scenario (Geometry A, S5, $N=4$, 60 s) and verify: trajectory file loads in PedPy, per-cell occupancy GT is sane, BLE RSSI tensors right shape, placeholder Bagging-LGBM forward pass returns one MAE per cell.

Gate to Phase 1: end-to-end run completes in < 5 min wall-clock on CPU. (Per IP-072 Phase B — the placeholder model is LGBM, not a dummy HAN. Cheaper to debug.)

Phase 1 — JuPedSim sandbox + PedPy validation + regime axis

1. For each (geometry, scenario, N): 5 random-seed replicates of 600 s simulated time using CFSM. Record trajectories.
2. Run PedPy over each trajectory: Voronoi density per cell at 1 Hz; fundamental diagram; continuity-equation residual (this is now a first-class output, not a derived quantity).
3. Validation gate, double sanity check:
   • Fundamental diagram against Adrian et al. 2024 empirical band and Kladek-Weidmann speed-density overlay (wolinski2014_f409 F-metric).
   • Report Cordes Intrusion / Avoidance dimensionless numbers (Dimensionless numbers reveal distinct regimes in the structure and dynamics of pedestrian crowds ↗) per scenario — regime declaration explicit, not implicit.
   • Derive a spatially-smoothed cell-occupancy projection as the surrogate that maps onto BLE-counted occupancy semantics (Risk #5 mitigation).
4. Add AVM replicates as ablation axis on locomotion model.
5. New: log the continuity residual of the trajectory-derived Voronoi density at each window — this is the ground-truth signal the model's continuity-residual penalty is supposed to regularise toward. Without this Phase 1 deliverable, Phase 5's λ-sweep has nothing to compare against.

Gate to Phase 2: PedPy validation passes; Cordes regime numbers reported; per-cell density timeseries (Voronoi + smoothed-projection variants) exported as parquet; continuity-residual tensor exported per-window per-cell.

Phase 2 — BLE RSSI synthesis

Log-distance path loss with log-normal shadowing: $$\text{RSSI}(d) = \text{RSSI}0 - 10n \log{10}(d / d_0) + X_\sigma$$ with $n = 2.7$, $\sigma = 4,\text{dB}$, plus 6 dB per concrete wall. Phone-penetration mask: only carriers emit.

Effective vs nominal penetration — per Accurate occupancy estimation with WiFi and bluetooth/BLE packet capture ↗, report both nominal (sweep value) and effective (post-MAC-randomisation, multi-device-collapsed) penetration on every Fig 1 curve.

Sensitivity ablations to flush before Phase 3:

• σ-sweep ∈ {3, 4, 6} dB to bracket Memoryless Techniques and Wireless Technologies for Indoor Localization With the Internet of Things ↗ envelope. Headline runs stay at σ=4 dB.
• Crowd-density-dependent extra-loss term — +1 dB per 0.5 ped/m² intersecting agent-to-observer LoS.
• Motley-Keenan / COST-231 saturating-multi-wall (~3.4 dB first wall + ~1.9 dB per additional) vs linear 6 dB/wall, on Geometry B (deeper NLoS).

Gate to Phase 3: RSSI distributions match published indoor BLE measurements; σ-sweep ablation logged; Motley-Keenan ablation logged for Geometry B.

Phase 3 — Sionna RT channel synthesis (fidelity spot-check, scope reduced)

The headline λ-sweep result uses QuaDRiGa (fast-tier). Sionna RT runs only on Geometry A, S3, N=8 as a fidelity spot-check.

1. Build Mitsuba scenes from Geometry A.
2. Place TX/RX antenna geometries matching planned EXP-001 hardware (Intel AX210 pattern).
3. Step through trajectories at 10 Hz, place dielectric boxes at agent positions, run Sionna RT, extract per-subcarrier $H(f, t)$.
4. Aggregate to per-window CSI features matching the :CSIAP schema.
5. Three spot-checks (each pass / fail logged):
   • Single-agent LoS crossing — visible amplitude dip + phase rotation. Otherwise dielectric-box model is broken.
   • Anechoic vs furnished delta — vs empirical contrast in A CSI Dataset for Wireless Human Sensing on 80 MHz Wi-Fi Channels ↗. If multipath floor under-represented → flag NIST-Q-D hybrid as needed (per Integrated Sensing and Communication: Enabling Techniques, Applications, Tools and Data Sets, Standardization, and Future Directions ↗).
   • Dielectric-box vs 3-layer human ablation — re-run with 3-layer dielectric; report per-cell MAE delta vs the cheap box model. Bounds body-modelling fidelity error.

Gate to Phase 4: synthetic CSI for Geometry A, S3, N=8, 5 replicates × 600 s built and cached. Three spot-checks logged. The Sionna result either confirms the QuaDRiGa λ-sweep trend (headline holds) or contradicts it (headline downgraded to "fast-tier upper bound"; investigation required).

Phase 4 — Model zoo + λ-sweep harness

1. Per 1-second window, materialise the input tensor for every model in the zoo. Six different input shapes, one shared (trajectory, csi, ble, regime, continuity_residual_target) parquet:
   • Bagging-LGBM: flat feature vector per cell
   • CNN+LSTM: per-link CSI 2-D tensor + BLE vector
   • Concat-MLP: same as CNN+LSTM, flattened
   • PINN: same as CNN+LSTM, plus the cell-graph adjacency for the residual computation
   • HAN-GAT: full heterogeneous HeteroData graph (per Heterogeneous Graph Attention Network ↗)
   • GCN: same graph as HAN-GAT, single node type, uniform aggregation
   • TKL: kernel-space features per Device-free occupancy detection and crowd counting in smart buildings with WiFi-enabled IoT ↗ / FreeCount: Device-Free Crowd Counting with Commodity WiFi ↗
   • BLE-only LR: BLE count per cell
   • CSI-only CNN: per-link CSI features, no BLE
2. Loss schema (the experiment's centrepiece): $$ \mathcal{L} = \mathrm{MSE}(\hat{n}, n) ;+; \lambda \cdot \mathrm{MSE}(\partial \hat\rho/\partial t + \nabla \cdot (\hat\rho \mathbf{\hat v}), 0) $$ with $\lambda \in {0.0, 0.01, 0.1, 0.3, 1.0}$. The first term is the standard supervised loss; the second is the continuity-residual penalty. Every model in the zoo trains with this loss schema, with $\lambda=0$ as the no-penalty control.
3. Train / val / test split by replicate seed (avoids temporal leakage). Same split across the whole matrix.
4. Persist per-window continuity-residual + per-cell MAE per (model, λ, scenario, geometry, replicate).

Phase 5 — The λ-sweep × model-zoo matrix (headline result)

Baseline table — 9 models × 5 λ-values × 5 scenarios × 2 geometries × 5 replicates = 2 250 training runs on QuaDRiGa fast tier.

Three orthogonal axes the matrix answers:

1. Architecture axis (column λ=0): which model wins without the physics prior? Hypothesis (from mondal2023): Bagging-LGBM ≈ PINN > CNN+LSTM > HAN-GAT > GCN > concat-MLP, with the deeper models hurt by data scarcity.
2. Penalty axis (row Bagging-LGBM, or any fixed architecture): does $\lambda > 0$ reduce MAE? Hypothesis: yes, monotonically up to some $\lambda^$ then degradation. The location of $\lambda^$ is itself a thesis-figure ("optimal continuity weight is data-driven, not arbitrary").
3. Interaction axis (full matrix): is the penalty's contribution separable from the architecture? Hypothesis: the penalty lift is consistent in sign across all architectures, with bounded variance in magnitude. This is L4½.5 — the empirical separation claim.

Cross-domain mini-ablation (Lever 2 reframed): train on Geometry A only, evaluate on Geometry B held-out. Headline: the penalty $\lambda^*$ that wins on A should still help on B (mechanism transfers). Architecture-specific parameters require refit (architecture does not transfer). This is L4½.6.

Per-regime conditional analysis (synthesis-2026-05-19-density-regime-axis): every cell in the 2 250-run matrix is also broken out per regime label (free_flow / transition / stop_and_go / jammed_turbulent). Penalty lift is expected to concentrate in transition and stop_and_go per the synthesis brief.

Phase 6 — Physics-residual analysis + falsification (replaces attention-weight analysis)

The attention-weight story (old Phase 6) is removed — it was bait for a thesis chapter the redesign no longer requires. The new Phase 6 answers the falsification question for L4½.5:

1. For each (model, $\lambda > 0$) pair, compute the achieved continuity residual on test windows. Expectation: $\lambda > 0$ models have lower residuals than $\lambda = 0$ models, with the residual decreasing monotonically with $\lambda$ up to $\lambda^*$.
2. Correlation analysis: per-window achieved residual vs per-window MAE. The penalty's mechanism is "reducing residual ⇒ reducing MAE" — if this correlation is weak, the penalty is doing something other than physics-prior regularisation, and L4½.5 must be qualified.
3. Counter-test: train a model with a random conservation law (replace $\nabla \cdot (\rho \mathbf{v})$ with $\nabla \cdot (\rho \cdot \mathbf{r})$, $\mathbf{r}$ a random vector field with the same magnitude statistics). If random-law MAE matches continuity-law MAE, the penalty is acting as generic regularisation, not physics prior. This is the load-bearing falsification test.
4. Visualisation (candidate thesis figures): per-cell residual heatmaps overlaid on Geometry A and B floorplans for $\lambda = 0$ vs $\lambda^*$. Visual evidence that the penalty redistributes residual physically (concentrated at boundaries / doors, dispersed at cell interiors).

Expected Outputs

• experiments/exp006/geometry/{a_seminar,b_corridor}/{geometry.json,scene.xml} — Shapely + Mitsuba scenes
• experiments/exp006/trajectories/<scenario>/<seed>/positions.parquet — JuPedSim outputs (Git LFS)
• experiments/exp006/csi_quadriga/, experiments/exp006/csi_sionna/ — channel-model outputs (parquet, Git LFS)
• experiments/exp006/ble_rssi/<scenario>/<seed>/rssi.parquet — RSSI tensors with nominal_pen + effective_pen columns
• experiments/exp006/density/<scenario>/<seed>/density.parquet — Voronoi + smoothed-projection + continuity-residual target
• experiments/exp006/graphs/<scenario>/<seed>/snapshots.pt — PyG HeteroData per window
• experiments/exp006/models/<model_name>/<lambda>/<run_id>.pt — checkpoints per (model, λ)
• experiments/exp006/results/<model>_<lambda>_<scenario>_<geometry>.parquet — per-window MAE + achieved residual + per-regime breakdown
• experiments/exp006/figures/ — λ-sweep curves, residual heatmaps, fundamental diagrams, cross-architecture penalty-lift comparison

Analysis Plan

Primary metrics

1. Per-cell occupancy MAE averaged across cells and windows — per (model, λ, scenario, geometry).
2. Penalty lift $\Delta\mathrm{MAE}(\lambda) = \mathrm{MAE}(\lambda = 0) - \mathrm{MAE}(\lambda)$ — the headline ablation quantity.
3. Cross-architecture penalty consistency — variance of $\Delta\mathrm{MAE}(\lambda^*)$ across the model zoo. Small variance ⇒ mechanism is architecture-independent (L4½.5).
4. Optimal $\lambda^*$ across models — does every model peak at the same $\lambda$, or does deeper-architecture / more-data prefer larger λ?
5. Cross-geometry generalisation gap of $\lambda^*$ — $|\lambda^_B - \lambda^_A|$ trained on A only. Small gap ⇒ penalty weight transfers (L4½.6).
6. Regime-conditional penalty lift — penalty $\Delta\mathrm{MAE}$ per regime label. Expected: largest in transition / stop_and_go.
7. Achieved residual ↔ MAE correlation — per-window correlation; if weak, falsifies the mechanism story.
8. Random-law counter-test MAE — vs continuity-law MAE. Equal ⇒ falsifies physics-prior interpretation.

Key figures

• Fig 1. λ-sweep curves: MAE vs λ for all 9 models, Geometry A, S3, N=8. The headline figure. Replaces the old "MAE vs phone-penetration" headline.
• Fig 2. Cross-architecture penalty lift bar chart: $\Delta\mathrm{MAE}(\lambda^*)$ per model. Visual evidence that the lift is consistent in sign across the zoo (L4½.5).
• Fig 3. Cross-geometry transfer: $\lambda^$ on A vs $\lambda^$ on B per model. Visual evidence of mechanism transfer (L4½.6).
• Fig 4. Achieved-residual heatmap overlay: $\lambda = 0$ vs $\lambda^*$ on Geometry A and B floorplans. The "physics figure".
• Fig 5. Regime-conditional penalty lift: $\Delta\mathrm{MAE}(\lambda^*)$ per regime label. Connects to synthesis-2026-05-19-density-regime-axis.
• Fig 6. Random-law counter-test: continuity-law vs random-law MAE per model. The falsification figure.
• Fig 7. Fundamental diagram from PedPy vs Adrian-2024 — sanity for the simulation layer.

Success Criteria

• End-to-end pipeline runs Geometry A, S3, $N=8$, 5 replicates × 600 s in < 8 h wall-clock on CPU + 1 GPU (QuaDRiGa fast tier across the full model zoo; Sionna spot-check is the budget headroom)
• L4½.4 evidence: penalty lift $\Delta\mathrm{MAE}(\lambda^*) \geq 0.15$ ped/cell on at least 5 of the 9 models in the zoo at S3, Geometry A
• L4½.5 evidence: penalty-lift sign is consistent (positive) across at least 7 of the 9 models; standard deviation of $\Delta\mathrm{MAE}(\lambda^*)$ across models is bounded ($\leq$ 50 % of the mean lift)
• L4½.6 evidence: $|\lambda^_B - \lambda^_A|$ stays within an order of magnitude for the top 3 models by Geometry-A performance
• Mechanism falsification passes: continuity-law penalty produces a lower per-window achieved residual than random-law penalty, and that residual reduction correlates with MAE reduction ($\rho > 0.3$, both Pearson and Spearman)
• mondal2023 replication in simulation: Bagging-LGBM at $\lambda = 0$ achieves MAE within 1.5× of the best deep model at $\lambda = 0$ — confirms the briefing's data-scale prediction
• PedPy fundamental-diagram residual stays within published empirical band
• All 9 models reach Phase 5 without OOM or crash on the QuaDRiGa fast tier

Risks and Mitigations

1. Synthetic-CSI fidelity (unchanged from old EXP-006). Mitigation: report QuaDRiGa headline; use Sionna RT only for the one spot-check; do not market simulation MAEs as predictive of hardware MAEs; thesis claim is "mechanism validates", not "performance generalises". wang2026_2758 3–5 y horizon framing intact.
2. GPU envelope. 2 250 training runs on QuaDRiGa is feasible on a single mid-range GPU over ~1–2 weeks. The Sionna spot-check is the only Sionna RT run. Mitigation: Sionna scope is one (geometry, scenario, N) — explicitly de-scoped from full sweep.
3. Static-vs-dynamic graph topology. Mitigation: fix topology as static for this experiment; flag dynamic-edge-weight extension as follow-up (now lower priority — graph-attention is no longer the headline).
4. Training-data volume. Deeper models (HAN-GAT, CNN+LSTM) borderline on ~3 k labelled windows per scenario. Mitigation: this is expected and is part of the headline finding — mondal2023 predicts the deeper models will be hurt; the matrix should show it rather than hide it. Bagging-LGBM does not have this risk.
5. PedPy-derived ground truth ≠ EXP-001 BLE-counted ground truth. Mitigation: spatially-smoothed cell-occupancy projection of the same trajectories (Risk #5 of old EXP-006 design; mitigation retained verbatim).
6. Attention-as-explanation rebuttal. No longer applicable — attention-weight analysis is removed from Phase 6. Replaced by the more-defensible physics-residual analysis + random-law counter-test.
7. Channel-model under-modelling stack (specular RT, dielectric-box humans, per-frame retracing). Mitigation: unchanged from the old EXP-006 — three spot-checks in Phase 3, dielectric-box vs 3-layer ablation bounds the body-model error.
8. Multi-wall path-loss linearity. Mitigation: Phase 2 Motley-Keenan ablation; report whichever is more conservative.
9. Mechanism falsification (new for this design). The random-law counter-test in Phase 6 is the headline defense — if it fails (random law matches continuity law), L4½.5 collapses and the chapter has to be reframed. Mitigation: falsification result is reported regardless of sign; this is intentionally a make-or-break test for the central thesis claim.

Related Work in Vault

Physics-informed learning + continuity equation (the headline lineage)

• di2023_285b ↗ — PINN survey with continuity-residual weights $\lambda \in [0.01, 1.0]$; EXP-006's λ-sweep range comes from here.
• hughes2002_57b4 ↗ — pedestrian-flow continuity-equation formulation; the prior the penalty enforces.
• maury2018_d24a ↗ — Seyfried / Boltes fundamental-diagram references; continuity-equation framing for indoor crowds.
• huang2009_292f — numerical algorithms for solving the continuity PDE; informs the discretisation in Phase 1.
• zabin2026_a20c ↗ — closest published physics-aware CSI adaptation work; does not use the continuity equation explicitly (the gap EXP-006 fills).

Data-scale grounding (why architecture is not the headline)

• mondal2023_7f7a ↗ — Bagging-LGBM MAE 0.26 beats CNN+LSTM MAE 0.50 at this task scale. The reason the model zoo includes LGBM and the reason the headline is the penalty not the network.
• yang2023_a34a ↗ — multimodal-fusion taxonomy + reference library for the CNN+LSTM baseline.
• choi2022_17c2 ↗ — LGBM crowd-counting reference.

Cross-domain CSI (the non-GAT cross-domain answer)

• zou2018_1590 ↗ — TKL for cross-domain CSI; included in the model zoo.
• zou2017_9087 ↗ — same TKL formulation, earlier reference.
• chen2023_5cbd ↗ — five-axis DG taxonomy; positions EXP-006's penalty-transfer claim as a new sixth axis.
• wang2026_2758 ↗ — 3–5 y sim-to-real horizon; the framing anchor for Risk #1.
• hou2023_bf83 ↗ — real-to-real cross-domain lower bound (even matched domains need few-shot adaptation).

Channel synthesis (Sionna spot-check + QuaDRiGa headline)

• hoydis2023_7aa4 — Sionna RT paper.
• sun2026_2f5e ↗ — RT runtime-budget critique.
• wang2022_2397 ↗ — NIST Q-D hybrid specular + diffuse as an upgrade path.
• meneghello2023_0a93 ↗ — empirical anechoic-vs-furnished yardstick for Phase 3 spot-check.

Pedestrian-dynamics ground truth + regime axis

• duives2013_3924 ↗ — CFSM/AVM lineage.
• zhong2022_7cb2 ↗ — fundamental-diagram validation as the gate.
• liddle2022_ee2c ↗ — Voronoi density as ground-truth estimator.
• cordes2024_05a8 ↗ — Cordes Intrusion / Avoidance regime declarations.
• wolinski2014_f409 ↗ — F-metric for fundamental-diagram validation.

BLE channel

• demrozi2021_bf55 ↗ — counting RMSE saturation at ~5 occupants.
• longo2019_b72f ↗ — effective-vs-nominal penetration drift.
• sadowski2020_f2ba ↗ — empirical n ∈ [1.8, 3.5], σ ∈ [3, 7] dB envelope.

Particle-filter / data-assimilation (adjacent fusion baselines parked as future work)

• malleson2020_7b38 ↗ — PF for crowd state; could be added as a 10th baseline in a follow-up.
• ghorbani2023_c065 ↗ — Kalman / PF for crowd; same.

Heterogeneous GAT (now one baseline among many)

• wang2019_84a9 ↗ — HAN, retained as one baseline row in Phase 5.
• chen2024_b58b ↗ — cross-domain CSI via dynamic-topology GNN; retained as adjacent precedent but not load-bearing.

Still-needed

• Veličković et al. 2018 GAT (10.48550/arXiv.1710.10903) — only needed if Ch7 evaluation needs a GAT primary citation. No longer load-bearing for the chapter; can ship without if necessary.

Thesis pillars

• 2026-04-13 - BLE-Calibrated CSI Crowd Sensing — core direction; the continuity-equation regularisation comes from §"Physics-Informed Direction".
• 2026-05-10 - GAT for BLE-Calibrated CSI Crowd Sensing — superseded as a chapter-slot decision by 2026-05-19 - GAT thesis deployment strategy.
• 2026-05-19 - GAT thesis deployment strategy — the redesign rationale lives here.
• synthesis-2026-05-19-density-regime-axis — regime-conditional analysis carries into Phase 5.

Cross-experiment links

• EXP-001 — provides the hardware-side ground-truth definition.
• EXP-005 — directly comparable real-hardware fusion experiment.

Dependencies

• Compute: one NVIDIA GPU with ≥ 16 GB VRAM (Sionna RT spot-check + deep-model training in the zoo). CPU-only fallback: drop Sionna spot-check + skip CNN+LSTM / HAN-GAT / GCN rows; LGBM / TKL / PINN-lite still run the full λ-sweep on CPU.
• Software: pip install jupedsim pedpy sionna torch torch-geometric scikit-learn lightgbm. PyG ≥ 2.6.
• No vault dependencies — does not require FIIT floorplan, real BLE infrastructure, or EXP-001 through EXP-005 completion.

Notes

The thesis-quality narrative this experiment unlocks is:

"Before deploying the BLE calibration campaign on real CSI infrastructure (EXP-001–005), we validated in simulation that adding a continuity-residual penalty to the supervised loss measurably reduces per-cell occupancy MAE across a 9-model zoo — and that the lift is empirically separable from the architecture choice. The mechanism (Ch3 L3.4 / Ch5 L5.6), not the architecture (Ch4-CSI L4.5), is the contribution. Synthetic-CSI fidelity is a known limitation: the magnitude of the simulation-side lift is not predictive of the hardware-side lift. What is predictive is the architectural commitment — the continuity-residual mechanism transfers; the model-specific parameters do not."

The Geometry-A vs Geometry-B contrast remains a deliberate down-payment on Lever 2 (cross-site transfer), now reframed: the optimal penalty weight $\lambda^*$ transfers, not the architecture-specific parameters.

When the FIIT floorplan eventually lands (IP-040 + IP-062 tooling), this experiment's geometry factories swap for the real geometry with one function-call change — pipeline is geometry-agnostic.