examples/research-sota/05-cross-room-reid/README.md
Whether the same person can be identified across two different rooms with WiFi CSI. Three-tick arc: R3 baseline → R3.1 (architecture error negative) → R3.2 (corrected architecture, structurally validated).
| Script | Thread | Headline | State |
|---|---|---|---|
r3_crossroom_reid.py | R3 | MERIDIAN env-centroid subtraction recovers cross-room re-ID to 100% (synthetic 128-dim embedding setup) | POSITIVE |
r3_1_physics_informed_env.py | R3.1 | Physics-informed env subtraction at raw-CSI level FAILS (10% = chance). Even labelled MERIDIAN at raw CSI = 10%. Identifies architecture error: position-variance dominates at raw level. | NEGATIVE (architecture-error) |
r3_2_embedding_physics_env.py | R3.2 | Embedding-level physics-informed env: 20% = matches labelled oracle with zero labels. Architecture correct; per-subject signal weak in synthetic AETHER stand-in. | STRUCTURALLY VALIDATED |
R3.1 was the loop's first architecture-error negative:
| Kind | Example | Path forward |
|---|---|---|
| Missing-tool (revisitable) | R12 → R12 PABS | Tool arrives later; approach works |
| Physics-floor (permanent) | R13 contactless BP | Hard wall; no tool changes this |
| Architecture-error (correctable) | R3.1 (here) | Wrong application level; corrected sketch explicit |
raw CSI → AETHER embedding (position-invariant, ADR-024)
→ physics-informed env subtraction (uses R6.1 forward operator)
→ cosine K-NN against per-installation gallery
The physics-informed env prediction must operate at embedding level, not raw level. AETHER does position-invariance; predicted-env removes the remaining room-shift component.
Cross-room re-ID is a powerful biometric. The loop's privacy framework enforces:
ADR-107 Layer 5 (per-installation embedding-space rotation key) enforces these technically.
docs/research/sota-2026-05-22/R3-*.md, R3_1-*.md, R3_2-*.md01-physics-floor/ (R6.1 forward operator)