c-coverage-estimator-factorial — session 01KWYBZAN8MVDV6035949Z0ASD (2026-07-07)
Verdict: the estimator fix works — the wedged agents that raw-dominated 50% / 42% of two walks' footfall are bounded to <1% — and with it the honest picture emerges: the anchor-layout effect is dominated by a placement×anchor interaction, so no single "dispersed vs hub" verdict is valid on this floor. Criteria 1/3 (C1); C2's interaction-gate correctly fires (interaction CI excludes 0), which renders the anchor main effect inadmissible and blocks C3 — the ANOVA-before-pairwise discipline working as designed. Load-bearing caveat: the interaction's magnitude is partly engineered — placements A/B were chosen as through-hub-room vs avoiding-it precisely to stress the hub layout, so +45.8 pp is an upper bound on "adversarially-chosen routing," not a general placement-policy effect.
What we ran
2×2×2 factorial on resplan-12439's own geometry: {anchor: dispersed/hub} × {furniture: off/on} × {placement: A east→west / B north→south}, paired by seed. 20 walk-notebook (JuPedSim) runs (placement A/B × 10 seeds), all gate-passed; anchor + furniture applied post-hoc over each furniture-agnostic walk → 80 analysis cells. Anchors from the two staged experiments (dispersed = …-multiroom, hub = …-topo-anchors). Reduction coverage_factorial.py (unit = seed, n=10).
What we found
- Bounded-influence estimator works (C1 ✓) — the campaign's core fix. Per-agent dwell cap (20 frames) + a wedged-agent ≤10%-share backstop. On resplan-12439 (the floor whose sub-60 cm channels wedge agents, per c-resplan-egress) the wedge auto-flag fired on 2 of 20 walks, where a single wedged agent raw-dominated 50.2% and 41.7% of that walk's footfall respectively — capped to max 0.7%. Legitimate transit agents (max share 12.2%, near the ~1/14 natural per-agent level) are left untouched — the cap is scoped to wedged agents only, so this is bounded-influence for the wedge artifact, not a strong bound on every agent, and it rests on the wedge classifier (near-stationary bounding-box < 0.6 m). This is the class of contamination that made c-coverage-meets-crowds's 57% blind-spot headline 76%-one-agent; it cannot recur here.
- The anchor effect is interaction-dominated (C2 gate fires). The placement×anchor interaction is +45.8 pp (95% seed-bootstrap CI [+43.2, +48.4]) — far from 0. Hub anchors concentrated in living-0 cover the crowd well when it routes through living-0 (placement A) and poorly when it routes north–south (placement B); dispersed anchors hedge across rooms. Because the interaction CI excludes 0, the ANOVA-first discipline (criterion 2) says the anchor main effect must not be interpreted — and blocks C3. Honesty on the magnitude: A/B were hand-selected as through-hub vs around-hub routings, so an interaction between "routes through the hub anchors" and "hub vs dispersed" is close to tautological — +45.8 pp measures these two adversarial placements, not placement policy at large. The sign and existence of a routing dependence is the finding; the magnitude is an engineered upper bound.
- The main effect: statistically real but inadmissible under the gate (C3 not interpretable). Averaged over furniture+placement, dispersed shows +12.7 pp more blind-spot footfall than hub (95% CI [+10.3, +15.1], per-seed +7.6…+18.1) — the numeric test passes (excludes 0, ≥5 pp, all 10 seeds positive). But criterion 2's fired interaction gate makes it inadmissible, not failed: it averages over a factor (placement) that swings the effect by 46 pp, so it is not a geometry-only anchor effect. C3 is booked unmet because the gate blocks interpretation, not because the effect is null — a distinction the metrics.json preserves (
excludes_0=true).
(Method note, disclosed as a design deviation: the brief pre-registered "furniture fixed," but an initial run placed synthetic furniture at room centroids, which coincided with the per-room dispersed anchors and artificially blocked their LOS — dispersed-on read 78%. Furniture was repositioned to floor-interior points ≥250 cm from every anchor (an arbitrary threshold, an unregistered researcher degree-of-freedom) and the reduction re-run; the main effect fell 33→13 pp and the interaction held. The correction is honest and its direction — the effect shrank — argues against furniture-position p-hacking, but it is a deviation from the registered design, not a clean pass.)
What it means for the thesis chain
The deployment question ble-periodic-calibration / layout-topology-fingerprint-discriminability keep asking — dispersed or hub anchors? — has no geometry-only answer: the choice interacts strongly with crowd routing. Hub concentration is low-blind-spot only when the crowd flows through the hub room; dispersed placement hedges against unknown routing at a modest average cost. For IP-106 the deliverable is a warning, not a rule: anchor-placement A/B tests must control crowd routing, or the main effect is an artefact of which flow happened to be sampled. No hypothesis strength change (in-silico, one geometry). The bounded-influence estimator + wedge auto-flag should be back-ported to the 7-floor coverage corpus before any cross-floor coverage claim is reused, per the brief.
Honest scope
In-silico, one geometry (resplan-12439), LOS-visibility coverage proxy (not a propagation model), self-authored synthetic furniture as LOS occluders (affects sensing LOS, not navigation — the walk is furniture-agnostic). n=10 seeds, paired. Reduction + coverage_factorial.parquet (per seed×cell) + 3 figures (seed_trace, paired_diff with the ±5 pp band, fig_coverage_summary with seed-bootstrap CIs) under this session's artefacts/ prefix; 20 walks attached.