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

01KWWB270C80Y3RS9RRBY3A8D0

finished 2026-07-06 18:29:18.988403+00:00 → 2026-07-06 18:37:09.595601+00:00 · 18 runs · supervisor: claude-code

“Egress works where floors clear (16157: rho=0.96, exact p=0.011, super-linear growth) but sub-60cm walkable channels wedge-deadlock 2-6 agents under converging load on 1374/12439 (matching their QC warnings, non-monotone in load) — clearance right-censored, never averaged; 2/3 floors survive BH-FDR on monotonicity; criteria booked 1/4 by the letter with the mean-speed band broken exactly in congested cells (queueing signature). n=2 seeds: descriptive scope. Fix the wedge (hard QC gate or anti-wedge behaviour) before crowd-safety claims.”

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
Every run returns trajectory.html, footfall_heatmap.png, speed_vs_time.png, and a non-empty trajectory.parquet. yes
Kinematic gate on every run: n_outside_walkable == 0, mean_walk_speed_m_s within 0.8–1.6 m/s, peak_walk_speed_m_s ≤ 2.5. no
Full egress at intensity ≤ 1.0: egress.departed == spawned and turned_away == 0 on every floor. no
Egress dynamics respond to load: clearance grows super-linearly with intensity (Spearman ρ ≥ +0.6 pooled across seeds), steeper on the narrowest-exit floor. no

Synthesis

c-resplan-egress — first execution (18-run grid, deterministic supervisor)

What we tested

All 18 planned cells ran and gate-passed: 3 floors (resplan-1374, -16157, -12439) × intensity {0.5, 1.0, 2.0} × 2 seeds, 24 base agents spawning across the whole floor and heading for the west exit region. Criterion 1 (surfaces) was checked per run; criteria 2–4 from the reduction (egress_reduction.parquet, per-cell, pushed to this session's artefacts with metrics.json + figure + reduction source).

What we found

Where the floor clears, egress responds to load exactly as the brief predicted. resplan-16157 is the clean case: full egress at every intensity, clearance 303→876 frames from intensity 0.5→2.0, growth steepening with load (0.5→1.0: +34%; 1.0→2.0: +83% — the super-linear doorway-bottleneck signature), Spearman ρ = 0.96 (n = 6 points from 2 seeds × 3 intensities; exact permutation p = 0.011, one-sided as pre-declared). Where the floor doesn't clear, the cause is wedge deadlock, not throughput: resplan-1374 strands exactly 2 agents at intensity 1.0 on both seeds (22/24; its walkable-QC flagged 0.84 m² of sub-60 cm channels — one body passes, two deadlock), yet clears all 48 at intensity 2.0 on seed 1 in 527 frames — non-monotone, so the failures are stochastic wedging in narrow channels, the same failure family as the coverage campaign's stuck agent. resplan-12439 (1.07 m² narrow channels) strands 2–6 agents at intensity ≥ 1.0 in 3 of 4 cells. Clearance for non-clearing cells is right-censored at the 6000-frame horizon and never averaged; censored ranks enter ρ conservatively (12439: ρ = 0.89, exact p = 0.033; 1374: ρ = 0.64, p = 0.13). After BH-FDR across the 3 floors, 2 of 3 monotonicity tests survive at 0.05. turned_away = 0 and n_outside_walkable = 0 everywhere; peak speed ≤ 1.2 m/s everywhere. Criterion 2's mean-speed band (0.8–1.6 m/s) is broken in exactly the 4 congested cells (0.64–0.67 m/s) — run-long mean speed includes queue time, so under egress load the band violation is the congestion signature itself, not a kinematic fault; booked false per the letter.

Statistics scope (n = 2 seeds/cell): all comparisons are descriptive with exact tests; at this replication no cell-level significance below the enumeration floor was attainable, and none is claimed.

What it means for the thesis chain

The walker reproduces load-responsive egress with visible doorway bottlenecks on real multi-room floors (thesis/crowd-modeling), and a draining floor is confirmed as a strong non-stationary occupancy stressor for downstream CSI. The load-bearing negative is infrastructural: sub-60 cm walkable channels deadlock under converging load, exactly where the QC warning points (worst near (21.6, 0.3) on 1374). Before any crowd-safety claim on these floors: either the walkable-QC jam-risk warning becomes a hard pre-run gate, or the walker needs an anti-wedge behaviour (radius jitter / retreat-and-retry). A follow-on that varies exit-corridor width on 16157 (the floor that actually clears) would make the width-capacity claim testable; on 1374/12439 the wedge fix comes first.

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.

Criticism — c-resplan-egress / 01KWWB270C80Y3RS9RRBY3A8D0 (operator statistical audit, inline)

Session run by a deterministic operator-supervisor under the 2026-07-06 statistics contract; no LLM critic subagent was spawned (0 LLM tokens of the budget used). Contract-compliance notes:

  • Right-censoring handled per assumptions.md (censored cells reported as null + censored_at, excluded from means, conservative ranks in rho).
  • Exact permutation p at the true label universe; BH-FDR(3) declared across floors before computing; n=2 seeds scope capped to descriptive throughout.
  • The figure was inspected after rendering and corrected once (its first version averaged censored horizons into the accent line and used a glyph missing from Helvetica) — the inspect-after-render rule caught both.
  • Honest weaknesses: 2 seeds cannot separate wedge probability from load effect on 1374/12439; the "steeper on narrowest corridor" clause of C4 is unevaluable under censoring; duration 6000 may under-report clearance for slow-but-clearing cells (censoring conflates "wedged" with "slow" — a longer-horizon confirm run would separate them).

Attached runs

Run Gate Purpose Replay
G3JD19EC replay
DHPNKYCS replay
J38WARVW replay
44NWG82K replay
1JH8DBJV replay
YJG7ZT6H replay
WYN4FR1A replay
J321P0Q1 replay
1MJPN4PE replay
YZV911AN replay
BVENTBPE replay
3DZDWYY6 replay
V30VARVK replay
8DKXHT20 replay
M1XCZAFB replay
WSQBV8W0 replay
MZCBSC4S replay
MHTNYK3Y replay