The question: Now that CSI capture, drift, calibration, and multi-room transfer are mapped, what are the open questions that no current experiment or paper fully answers?
Five months from the October FIIT-library deployment, the lab has names for almost everything in the CSI pipeline — channel state information (the per-subcarrier amplitude and phase snapshots a WiFi card reports for every received packet), environmental drift, BLE-assisted transfer, model selection. What it does not have is answers to three questions that sit just outside every experiment currently running. This episode names them, ranks them by urgency, and asks which one breaks first when the building fills with students.
The deadline is fixed; the answers are not
At the meetings/2026-05-13-csi-experiments.md meeting, Jakub locked October as the FIIT-library deployment date: live data over summer, DP2 preliminary results in January. The date is not movable. What remained open: the data-transmission protocol (gRPC from Lukáš Adamík's bachelor thesis versus raw TCP), the calibration cadence, and whether an adjacent-room interference model even exists yet. experiments/EXP-001 BLE-Assisted CSI Ground Truth Collection.md handles ground-truth labelling, but the structural gaps above it are a different problem — they are not implementation choices, they are unanswered research questions with a hard ship date attached.
Q1 — How often must fine-tuning fire?
experiments/EXP-002 CSI Model Drift Measurement Over Time.md will tell us which model family — Random Forest, LSTM, or CNN — degrades slowest under drift. That is useful. What it will not tell us is the cadence at which BLE-triggered fine-tuning must fire to keep any of them above an acceptable accuracy floor. billah2021_69a2 puts a concrete number on why this matters: a comparable BLE-based occupancy system drops from 92.2 % to 70.3 % F1 over seven days without retraining. Seven days is not a long weekend — it is a normal gap between scheduled maintenance windows.
Before designing a retraining schedule, ask what "acceptable floor" means in a library context. Is 70 % F1 a fire alarm or a rounding error? The answer changes everything downstream.
Q2 — What happens when Room A's crowd leaks into Room B's signal?
experiments/EXP-004 Multi-Room Generalization with BLE Transfer.md targets more than 80 % of native accuracy after a short BLE-labelled calibration in a second room. The design assumes one room at a time. It does not model the case where occupancy in an adjacent room perturbs the target room's CSI — a physical reality in any building where walls are not RF-opaque.
jung2025_3f2e calls adjacent-room interference the single most critical unresolved issue in multi-zone deployments. The FIIT library is a multi-zone deployment. No current vault experiment touches this.
wang2026_2758 offers a cross-domain taxonomy that helps classify which failure mode this actually is — and the classification matters. If the interference is additive (a constant bias you could subtract), a simple correction layer might be enough. If it is multiplicative (a room-dependent scaling that shifts with crowd size), the entire transfer design needs rethinking before October.
Q3 — Does model complexity matter once drift correction is running?
The lab's architectural complexity story was partly collapsed when mondal2023_7f7a reported a Bagging-LGBM model achieving MAE 0.26 against a CNN+LSTM stack at MAE 0.50 on a crowd-counting task at the same scale the lab is targeting. The 2026-05-19 GAT reversal — a more complex architecture was planned, vault evidence said a simpler baseline already dominated, and the plan was dropped — is a cautionary instance of this pattern, not a resolution of it.
The deeper question is still open: does architectural choice matter at all once stage-four drift correction is running? guarino2026_e72c ↗ notes that roughly half of all public CSI datasets were collected in a single environment, making cross-environment complexity claims structurally untestable. huang2025_060d — the first benchmark with explicit multi-user identification, localisation, and activity-recognition splits — is sitting in the vault unengaged.
If Bagging-LGBM already beats the deep stack at this scale, what would have to be true about the FIIT library's crowd dynamics for a graph-based model to recover its advantage? That is not a rhetorical question; it is the experiment the lab has not designed yet.
Two experiments worth running before October
The interference probe. Place one router-and-receiver pair in Room B. In an adjacent Room A, cycle through 0, 5, 10, and 20 occupants without changing Room B's occupancy. Plot Room B's predicted count as Room A fills. If the prediction moves, you have quantified the interference signal. If it does not, you have evidence the wall attenuates enough to ignore. Either result is a result.
The cadence sweep. Train a model, freeze it, and measure accuracy at 1-day, 3-day, 7-day, and 14-day intervals without retraining. Then run BLE-triggered fine-tuning at each interval independently and measure the accuracy recovery. The gap between degraded and recovered accuracy at each cadence point tells you the cost of waiting. hou2023_bf83 provides a published analogue for the transfer step — worth reading before deciding what "recovery" should look like numerically.
One reading to start with
jung2025_3f2e is the single best entry point because it is the only paper that runs a CSI occupancy system in a multi-zone environment and names, specifically, what broke. Read it before designing any experiment that touches more than one room.
If the lab can instrument only one new experiment before October, the choice is stark: the adjacent-room interference case that no current design touches and that can silently corrupt every room's count simultaneously — or the multi-person disambiguation case that huang2025_060d already benchmarks and experiments/EXP-005 Multi-Person Activity Disambiguation.md has barely started? Which failure mode do you want to discover in a controlled test — and which one are you willing to discover live, with students in the building?