01KW022WMG3X4PF8PA4ZN697ET

Coverage meets crowds — does the placed-anchor footprint cover where the crowd goes?

A through-traffic crowd (14 agents, seed 0) walked each of 7 floors (sweep 01KW01VXFW00C1MCTDHSSNQ0FW); a first-order line-of-sight reduction (a cell is sensed when an anchor within 8 m has a wall-unobstructed segment to it; anchors from gis_experiment_layout) overlaid the time-integrated footfall and computed the fraction of footfall in sensing blind-spots.

C1 — met. All 7 walks pass the gate; sampled metrics confirm n_outside_walkable == 0 and mean speed ~1.19 m/s. Caveat: the load-bearing floor, resplan-12439, departed 13/14 (1 stuck) and carries wedge/jam QC warnings — not a perfectly clean walk.

C2 — met. Per-floor overlays + coverage_meets_crowds.parquet carry covered_area_frac, footfall_coverage, blindspot_footfall_frac, mean_coverage_depth_weighted; no NaN/Inf. (floor_area_m2 here is envelope-interior area, a coarser measure than the walk's walkable_qc.area_m2 — provenance differs; the blind-spot fraction is footfall-weighted and independent of it.)

C3 — met. Ranked blind-spot footfall: test-lab-synth 0% · 12419 1% · 16157 1% · 7421 5% · 1374 14% · 147440 22% · 12439 57%. Among the four-anchor dispersed floors it rises broadly with area (mean 14%) — consistent with an area/anchor-budget constraint. LOS-proxy framing held throughout: this ranks placements and exposes demand/coverage mismatch; it is not a measured detection rate.

C4 — reported (causal claim withheld). The 12439 hub-concentrated layout does not lower blind-spot footfall — it is the worst in the corpus (57%): its anchors sit on the central hub while the through-traffic crosses peripheral rooms/corridors. But at N=1 this is confounded — 12439 differs from the dispersed floors on four axes at once (6 anchors vs 4, larger area, the only furnished floor with 16 occupiable, and placement). The result is therefore associated with hub concentration, not demonstrably caused by it, and is specific to a through-traffic demand pattern; a hub-traversal crowd would likely invert it. The honest deliverable is the demand/coverage-mismatch diagnostic and the open hypothesis that count-optimised placement (hub) and demand-coverage-optimised placement diverge — to be isolated by a controlled anchor-count/placement sweep and confirmed radiometrically (exp-csi-static) / in hardware (IP-106).

Critic (severity: medium, did NOT agree with the original causal verdict — integrated)

Borderline criterion: C4. The 0.570 number is real but the synthesis originally asserted a single mechanism ("because hub anchors miss the through-traffic"). At N=1, resplan-12439 is confounded on ≥4 axes simultaneously — 6 anchors (not 4), larger area, the only furnished floor (16 occupiable, so furniture also occludes LOS), and placement strategy. The final synthesis now says associated with, names the confound, and reframes the divergence as an open hypothesis requiring a controlled sweep.

Other corrections integrated: (1) C1 marked true partly on gate_passed; the load-bearing 12439 floor actually departed 13/14 with wedge/jam QC warnings — now stated. (2) no-NaN scoped to the four coverage metrics (the walk metrics carry by-design NaNs out of scope). (3) Area-provenance mismatch named (coverage floor_area_m2 = envelope interior, e.g. test-lab 64 m² vs walk walkable_qc 15.7 m²).

Scope: N=7, single seed, single LOS range (8 m) and grid bin (0.25 m), near-free-flow (no congestion stress); LOS ≠ RF detection. The ranking and the LOS framing are sound; the single-cause attribution was the weak point and has been removed.

Addendum — post-hoc statistical re-analysis (2026-07-06, statistics skill, operator session)

Independent recomputation from the session's raw run artefacts (trajectory.parquet, floor_geometry.json, walk-notebook.metrics.json per floor) driven through the unmodified published reduction (monad_knowledge/notebooks/python/coverage_meets_crowds.py) under the house statistics contract (.claude/skills/statistics/). The headline numbers reproduce (blindspot_footfall_frac within 0.40 pp on every floor), but re-derivation surfaces a measurement artefact that changes the campaign's central ranking fact. Verdict direction on C4 ("associated with, not caused by") is upheld and reinforced; one ranking claim is overturned; four precisifications follow.

• [HIGH → overturns C3/C4 ranking] The 57% blind-spot at resplan-12439 is a stuck-agent artefact, not a placement signal. One agent (id 1) contributes 50.2% of the floor's 7,937 footfall frames — it is wedged within 0.1 m of its spawn from t = 1.0 s and never displaces more than 0.49 m across the full 200 s (239 m of in-place jitter path, tail position std 0.030 m), sitting in a cell with LOS to no anchor. Because the footfall estimator is a raw per-frame 2-D histogram with no dwell cap or per-agent weight normalization, the mean's unbounded influence function (gentle2020_1ba7 ch. 8.7; assumptions.md robustness row) lets that single frozen agent carry 76.3% of the blind-spot value. Excluding it: blindspot_footfall_frac 0.5695 → 0.1349, which moves 12439 from rank #1 to ~#3 in the corpus — tying resplan-1374 (0.138) and below resplan-147440 (0.218). The claim "worst in the corpus (57%)" (C3/C4, supervisor verdict) does not survive de-contamination. A systematic sweep of all 7 floors flags only 12439 (every other floor's busiest agent ≤ 11.3% of frames with 3.8–6.8 m tail std — ordinary walking). The floor's own walkable_qc had already emitted a wedge-risk warning (0.48 m doorway, core splits into 2 parts) and the run departed 13/14 — the mechanism was in hand at C1 and never propagated to the coverage metric. Correct phrasing: 12439's raw blind-spot fraction is dominated by one wedged agent; the de-contaminated value (~13.5%) is mid-pack, and the corpus ranking does not single 12439 out once the artefact is removed. Fix the estimator (bounded per-agent contribution / unique-cell weighting) before any 12439 conclusion.

• [MEDIUM] The "(mean 14%)" attributed to the dispersed floors is computed on the wrong set. mean(blindspot) over the six 4-anchor dispersed floors = 7.06%; over all seven (incl. the 6-anchor 12439) = 14.18%. The stated "(mean 14%)" is the all-seven mean — which is itself inflated by the artefact above — not the dispersed-only subset the sentence names (tebbs2006_698b ch. 9, internal-consistency check on a headline number). Report the dispersed-only 7% or drop the parenthetical.

• [MEDIUM] The area/anchor-budget trend is directionally real but not robustly significant at n = 6. The synthesis asserts "rises broadly with area" with no statistic, contrary to the house rule "estimate + CI + effect size, then p" (SKILL §1.5). Quantified over the six dispersed floors: Spearman ρ = 0.943, exact permutation p = 0.0167 (720 relabelings; the asymptotic p = 0.0048 is optimistic at this n — SKILL §2). But the significance hinges on two soft choices: Pearson on the same six gives r = 0.805, p = 0.053 (n.s.); and excluding test-lab-synth (a synthetic control with blindspot = 0 by construction, not a ResPlan draw) the five real floors give ρ = 0.900, exact p = 0.083 (n.s.). The trend is not strictly monotone either (147440, 312 m², 21.8% exceeds 1374, 389 m², 13.8%). Report it as a suggestive, unreplicated, floor-level (ecological) association at n = 6, not an established relationship (peck2008_2ba0 ch. 2: purposive sample licenses no generalization).

• [MEDIUM] No headline number carries an interval; the only available resampling unit shows overlapping "ranks." With seeds: [0] there is no seed-level replication; the sole legitimate unit is the 14 agents per floor. Agent-bootstrap 95% CIs (B = 2000): 12419 1.1% [0.0, 2.8]; 16157 1.1% [0.2, 2.4]; 7421 4.6% [0.0, 9.8]; 1374 13.8% [5.5, 22.5]; 147440 21.8% [14.4, 28.4]; 12439 57.0% [8.4, 82.0]. The 12439 interval spanning nearly the whole [0,1] (boot-mean 0.465 ≪ point 0.570, because agent 1 is absent in ≈35% of resamples) is a direct fingerprint of the artefact above; several adjacent floors (7421 vs 12419/16157) overlap and should not be read as cleanly separated. Every mean over runs must carry a labelled interval (figures/references/statistical-figures.md); the corrected summary figure is the required default.

• [DESIGN] Follow-up sizing (fix the estimator first). Using the clean per-floor agent bootstrap SDs (σ ≈ 3.5–4.3 pp) as the prior, a 15 pp dispersed-vs-hub effect needs only ~1–2 seeds/arm and a 5 pp effect ~10 seeds/arm (campaign-design.md §3, n = 2σ²(z₀.₈+z₀.₀₂₅)²/ES²). If the estimator is not fixed, the contaminated variance regime (σ ≈ 0.26 from 12439's bootstrap) inflates the same 15 pp target to ~47 seeds/arm — a ~40× cost for skipping step 1. Isolate the four confounds (6 vs 4 anchors, area, furniture, placement) with a controlled anchor-layout factorial on fixed geometry (devore2012_62c8 ch. 11), ≥8 seeds/cell (assumptions.md small-n floor), and pre-declare the comparison.

Figures: fig-stuck-agent-contamination.png (per-agent frame share for 12439 + full-vs-excl blind-spot across all 7 floors), fig-blindspot-summary-ci.png (corrected cross-floor ranking with agent-bootstrap CIs). Sources beside them in the audit directory.

Corpus citations: gentle2020_1ba7 (unbounded influence, ch. 8.7), wasserman2004_ea08 (bootstrap percentile CI, ch. 8; small-n honesty, ch. 8.3), diez2015_8380 (exact permutation, ch. 6), tebbs2006_698b (headline-number consistency, ch. 9), peck2008_2ba0 (scope of inference, ch. 2), devore2012_62c8 (factorial screening, ch. 11), campaign-design.md (paired power), and figures/references/statistical-figures.md (bare-point-estimate rule).