Campaign

Coverage factorial with a bounded-influence estimator — isolate what actually drives blind-spot footfall

c-coverage-estimator-factorial · walk-notebook

Archive snapshot, as of 18 h 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.

Sessions

Session	State	Runs	Synthesis	Criticism	Figures	Verdict
R8N2MR9W 2026-07-07T20:57	finished	20	✓	—	—	Committed-reduction reproduction of the canonical session 01KWYBZAN8MVDV6035949Z0ASD. 20 fresh walk-notebook runs (placement A/B × 10 seeds), all gate-passed; the committed coverage_factorial.py reproduces the sealed result byte-for-byte: bounded-influence max wedged-agent share 0.7% (C1 ✓, wedge auto-flag fired 2/20), placement×anchor interaction +45.8 pp [+43.2,+48.4] (gate fires → main effect inadmissible), anchor main effect +12.7 pp [+10.3,+15.1]. Confirms the reduction now in git + the GHCR image matches the working copy. Criteria 1/3, identical to canonical. No new science — reproduction check only; see 01KWYBZAN8 for the full synthesis + critic.
949Z0ASD 2026-07-07T13:23	finished	20	✓	✓	—	Estimator fix works — wedged agents that raw-dominated 50%/42% of two walks are bounded to <1% (max wedged share 0.7%; transit agents near the ~1/14 natural level untouched) — the bounded-influence footfall estimator + wedge auto-flag are the campaign's core deliverable and succeed. The anchor-layout effect is interaction-dominated: placement×anchor interaction +45.8 pp (95% CI [+43.2,+48.4]) excludes 0, so the interaction gate fires and the anchor main effect is inadmissible. Honesty: A/B placements were hand-selected as through-hub vs around-hub, so the interaction magnitude is engineered — the sign/existence of routing-dependence is the finding, the magnitude an upper bound. Anchor main effect +12.7 pp [+10.3,+15.1] (10/10 seeds positive, ≥5pp) is statistically real but inadmissible under the gate — booked unmet, not null (metrics preserves excludes_0=true). Criteria 1/3 (C1). No geometry-only dispersed-vs-hub answer on this floor; IP-106 anchor A/B tests must control routing. No hypothesis strength change; back-port estimator to 7-floor corpus. 20 walks gate-passed. (Design deviation disclosed: synthetic furniture repositioned ≥250cm off anchors after an initial centroid placement confounded dispersed; main effect fell 33→13pp, interaction held.)

Brief

Why this campaign exists

c-coverage-meets-crowds's headline blind-spot number (57% on resplan-12439) was 76% one physically wedged agent — the floor's own walkable-QC wedge warning fired and never reached the coverage stage — and four axes (anchor count, area, furniture, placement) were fully confounded across floors, at zero seed replication. The 2026-07-06 audit quantified the estimator's cost: the contaminated σ regime needs ~47 seeds/arm; the clean regime needs ~10 for a 5 pp effect at 80% power. Fix the estimator, then isolate the axes on one geometry (devore2012_62c8 ch. 11: interactions first).

Design

2×2×2 factorial on resplan-12439's own geometry: {anchor layout: dispersed/hub} × {furniture: on/off} × {placement policy: A/B}, geometry fixed, paired by seed across cells. 8 cells × 10 seeds = 80 runs.

Artefact contract (hard gate)

Bounded-influence footfall estimator wired into the reduction (step 1, non-negotiable); coverage_factorial.parquet (per seed × cell); factorial effect table with interaction + main-effect CIs; everything pushed to the session S3 prefix — the original never persisted coverage_meets_crowds.parquet despite all four criteria citing it.

Figure render request (mandatory)

seed_trace, paired_diff (dispersed − hub per seed).
Corrected fig_coverage_summary with seed-bootstrap CI error bars (the audit's fig-blindspot-summary-ci.py form).

Honest scope

One geometry by design — this is the isolation campaign; generalization is c-csi-cross-geometry-scaleout's pattern. The estimator fix must be back-ported and the 7-floor corpus re-run before any cross-floor coverage claim is trusted.