Asking the fleet what it is doing…
monad-knowledge Wi-Fi sensing lab · FIIT STU
Campaign session

01KW094BB5ZKAZ13TEDRM17SVP

finished 2026-06-25 20:56:47.717404+00:00 → 2026-06-25 20:57:36.236552+00:00 · 6 runs · supervisor: react-agent

“Denser-CSI + higher-occupancy did NOT meaningfully lift the CSI floor. On resplan-12439 with a NEW 10-Rx 'dense10' experiment + varying high occupancy (occ to 18, real arrival/departure variance), the CSI temporal-CV explains almost no count variance: 10-link per-link-ridge R2=0.09 (the only config above 0), 10-link scalar 0.00, 5-link both ~0/negative; nMAE ~0.23-0.26. Both levers help DIRECTIONALLY (10>5 links, per-link>scalar) but the absolute signal stays very weak. The earlier 0.16 was inflated by near-constant (seated-stay) occupancy. Conclusion: more links/occupancy is NOT the lever — the quasi-static ray-traced amplitude-CV feature is a fundamentally weak count predictor here. The real remaining lever is richer TEMPORAL CSI features (doppler micro-fading, off in these runs for frame alignment; or a learned spectro-temporal feature), or real CSI. Reinforces the arc: BLE device-counting is the load-bearing modality; CSI is a weak coarse fallback in the staleness-switch.”

Archive snapshot, as of 1 day ago — the run corpus is rebuilt once a day, so this page is not a live reading. The fleet panel is the live one; it refreshes every 30 s.

Success criteria

CriterionResolved
New 10-Rx dense experiment authored (multiroom+topo placements) + high-occupancy coupled runs at matched occupancy/seeds vs a 5-Rx control; varying-occupancy rerun for proper variance. yes
fusion_dense reduction reports CSI->count R2/nMAE/MAE by layout x feature (scalar-CV vs per-link ridge), 3-seed. yes
Floor-lift: NOT achieved meaningfully (best R2=0.09); denser links + per-link feature help only directionally; absolute CSI count signal stays weak. no
Honest framing: in-silico quasi-static CSI; the limit is the feature/signal, not link count; next lever = richer temporal CSI, not more anchors. yes

Synthesis

Denser-CSI + higher-occupancy floor-lift — does CSI quality rise with more links/crowd?

The staleness-switch's sparse-regime floor is set by CSI->count quality (~weak). This session tests the lever directly: a NEW 10-Rx dense experiment on resplan-12439 (csi-link-resplan-12439-dense10, combining the multiroom + topo placements), high occupancy, vs a 5-Rx multiroom control, 3 seeds each, with a per-link CSI feature (full per-link CV vector → ridge) vs the old scalar mean-CV. Reported as R^2 (count variance explained — scale-free), nMAE, MAE.

First run (seated-STAY, occ ~constant at 17): 10-link per-link R^2=0.16, nMAE 0.15 — looked promising, but occupancy was near-constant (seated agents stayed), so the low variance flattered both R^2 and nMAE.

Corrected run (seated arrive AND depart, occ varies 0->18): the result deflates honestly —

  • 10-link per-link-ridge: R^2=0.09, nMAE 0.23
  • 10-link scalar-CV: R^2=0.00
  • 5-link per-link / scalar: R^2 ~ -0.02 / -0.01

Verdict: the floor was NOT meaningfully lifted. Both levers help directionally — more links (10 > 5: R^2 0.00→0.09 per-link) and a richer feature (per-link > scalar: 0.00→0.09) each move it up, and they are the only configs above R^2=0 — but the absolute signal stays very weak: even 10 links + crowds to 18 + proper variance, the CSI temporal-CV explains <10% of count variance (~23% relative error). The synthetic, quasi-static ray-traced amplitude-CV is a fundamentally poor count predictor on these floors; adding receivers does not fix that.

What this means. The CSI floor cannot be raised by link count / occupancy alone — the bottleneck is the feature/signal, not the geometry. The real remaining levers are (1) richer temporal CSI structure — doppler micro-fading (deliberately off here for frame-axis alignment; flagged as the count-relevant carrier in c-csi-band-calibration), or a learned spectro-temporal feature instead of a single CV scalar/vector; and (2) real CSI (hardware temporal richness the quasi-static sim lacks, IP-106). This reinforces the whole arc's conclusion: BLE device-counting is the load-bearing modality; CSI is a weak coarse fallback, and the staleness-switch is the right way to use it — but its sparse-regime ceiling is set by a CSI signal that more anchors won't rescue.

Caveats: in-silico; quasi-static RT (no doppler); amplitude-CV feature only (not learned); 3 seeds; one floor. The doppler-on + learned-feature rerun is the clean next probe before concluding CSI is irredeemably weak.

Criticism adversarial review

Written by the campaign-critic subagent against the brief's success criteria — read it as the counter-position to the synthesis above.

Self-critical notes

Caught my own confound mid-experiment: the first (seated-STAY) run gave R^2=0.16 / nMAE 0.15, which I did NOT report as the result because occupancy was near-constant (low variance flatters both metrics). The varying-occupancy rerun (R^2=0.09) is the honest number. Remaining limits: (a) quasi-static RT with doppler OFF — the count-relevant temporal fine-structure (band-calibration finding) is excluded, so this is a lower bound on CSI's potential, not a ceiling; (b) the feature is a hand-built CV (scalar/per-link), not a learned spectro-temporal representation that might extract more; (c) 3 seeds, one floor, modelled BLE. The honest claim is bounded: the simple amplitude-CV feature on quasi-static RT CSI is a weak count predictor that more links/occupancy do not rescue — not CSI is irredeemably weak. Doppler-on + a learned feature is the test that would settle that.

Attached runs

Run Gate Purpose Replay
AQXRDDS7 dense-csi-floorlift-varying replay
W2PHTFWM dense-csi-floorlift-varying replay
X67Q6F87 dense-csi-floorlift-varying replay
RDZ30DZP dense-csi-floorlift-varying replay
4BAQW9Q9 dense-csi-floorlift-varying replay
V21FVNNM dense-csi-floorlift-varying replay