c-placement-oracle — Session 01KXDYYKNY7NYQH40MQBF585JX (Stage 0–2 mechanism pilot)

Executive summary

The floorplan-as-placement-oracle loop runs end-to-end, and its central claim survives its first contact with real geometry: on 6 topology-diverse floors, the crowd-count-optimal AP placement is a materially different placement from the coverage-optimal one — on every floor. Mean placement divergence (1 − corner Jaccard) is 0.61 (median 0.67, range 0.40–0.86, MAD 0.095), every floor clears the pre-declared min_effect_size of 0.15, and the coverage objective is exactly submodular (0.0 violations) so greedy carries its 1 − 1/e guarantee (attained 0.95–1.00 of the brute-force optimum on all 6). The go/no-go gate the diary set — does counting-optimal ever diverge from coverage-optimal, or is placement just an area/budget problem? — resolves GO.

This is a mechanism-existence pilot (n = 6 floors, seed 0, geometric LOS coverage + capped-depth counting proxy — NOT yet ray-traced CSI). It licenses the expensive stages; it does not replace them.

Abstract

Coupling the IP-107 persona agenda engine (walk-notebook) to the Pillar-4b submodular placement reduction (coverage_submodular), we generated five-persona crowds (worker / patron / visitor / through-traffic / meeter; door → room → dwell → exit) on 6 floors spanning a designed synthetic lab and small-to-large ResPlan apartments, then computed per floor the footfall-weighted coverage objective, a capped-depth counting proxy, and the Pareto trade-off between them. Coverage is submodular and greedy is near-optimal everywhere; the counting- and coverage-optimal placements diverge on all 6 floors, with the counting optimum concentrating on the footfall hub/corridor and the coverage optimum spreading to the geometric extent. The propagation-real (ray-traced I(C;Φ)) confirmation and the powered 56-floor corpus are deferred to the CI stages.

Plain-language method

We dropped five kinds of simulated people onto each floor — some sit and work, some browse and read, some just pass through, some arrive as a group and meet — and let them walk door-to-door through the real wall layout. We recorded where they went (a footfall heat surface). Then, for each floor, we asked two questions of a grid of candidate AP sites: (a) where do k=4 APs go to cover the most of where people are, and (b) where do they go to best see the count of moving people? "Best coverage" is a set-cover problem with a proof (greedy is within 63% of optimal, guaranteed); "best counting" is not, so we treat it as a heuristic and check it against the exhaustive best. Then we measured how much those two answers differ.

Per-criterion results

• C1 — corpus-wide crowds (Stage 0–1). Partial (pilot). 6/56 grid floors produced clean walk-notebook runs: five personas present, replayable trajectory.html, gate_passed on all. The walkable-QC correctly flagged marginal geometry on resplan-1374 (a two-abreast jam channel) — a valid run, geometry noted. The full 56-floor × 3-seed grid is deferred to /campaign-systematic.
• C2 — cheap Pareto + divergence (Stage 2). Resolved (mechanism). Submodularity violation 0.0 on all 6; greedy/optimal 0.95–1.00, beats 1 − 1/e everywhere; counting- vs coverage-optimal Jaccard 0.14–0.60 (divergence 0.40–0.86), objectives differ on 6/6. The powered cross-floor distribution claim needs n = 56 (CI); the mechanism is decisive at pilot scale.
• C3 — ray-traced I(C;Φ) kernel (Stage 3). Deferred. Not attempted this session; the placement_oracle reduction that computes it is built + selftested (measured-MI + surrogate + 3-variant selection), ready for the Sionna-RT verify floors on CI.
• C4 — surrogate + verify (Stage 4). Deferred. Gated behind C3.
• C5 — three deliverables + 5-panel figure (Stage 5). Partial. The Stage-2 placement figures rendered per floor (coverage-opt vs counting-opt + Pareto); the full operational/experimental/both placement_result needs the ray-traced I(C;Φ) from C3.
• C6 — framing discipline. Held. Every number above is self-authored-crowd + geometric-proxy evidence; it ranks placements and is a hypothesis generator until a real AX210/Pi5 floor (IP-106 / IP-112) confirms the ranking. No hardware detection-rate claim.

Real-life meaning

If you are deploying Wi-Fi to both run a building and count its crowds, this says the two goals pull the APs to different places — you cannot get the count for free from a coverage-optimal rollout, and on a corridor-heavy floor (e.g. resplan-12419, Jaccard 0.14) the two layouts barely overlap. That is the deployment gap the campaign exists to quantify, and it is real at pilot scale. It also says which floor is worth instrumenting with real hardware first: the high-divergence floors, where the placement choice matters most.

Caveats

• n = 6, seed 0. A mechanism pilot, not the powered n = 56 claim; the cross-floor divergence distribution and its CI are a CI-stage deliverable.
• Geometric proxies. Coverage is first-order LOS; counting is a capped-depth footfall proxy — not ray-traced I(C;Φ). The propagation-real confirmation (Stage 3) is exactly where a self-confirming geometric surrogate could be wrong; that is the point of the verify stage.
• Sim-authored throughout. The standing project warning applies: a clean sim ranking means nothing until one real-hardware floor confirms it.

Critic audit (inline, tiny-corpus)

Claim audit. The headline (counting-opt ≠ coverage-opt on 6/6, mean divergence 0.61) is directly supported by coverage_submodular.parquet (pushed to the session S3 prefix, auditable) and the per-floor Pareto figures — the resplan-12419 figure visually confirms the counting optimum sitting on the footfall hub while the coverage optimum spreads. The submodularity/greedy-bound claim is exact-arithmetic, not estimated. No overreach in the "GO" verdict for the go/no-go gate.

Weak claims flagged + how the synthesis already handles them. (1) The counting objective here is a geometric capped-depth proxy, not ray-traced I(C;Φ) — the synthesis states this in C2/C3/caveats and does not claim the propagation-real result. Good. (2) n=6/seed-0 is below the declared powered n=56 — the synthesis scopes every claim as "mechanism pilot" and defers the powered distribution + CI to CI; it does NOT report a cross-floor CI it cannot support. Good. (3) "divergence tracks corridor-heaviness" is asserted qualitatively from one floor (12419) — correctly hedged as an illustrative case, not a fitted trend (that needs the 56-floor corpus + a topology covariate).

Honest scope. This session resolves C2 (mechanism) and C6 (framing) only; C1 is pilot-partial, C3–C5 deferred. The synthesis's criteria_resolved encoding reflects exactly this. No high-severity disagreement with the verdict.

Severity: low. The synthesis is appropriately sharp where the evidence is exact (submodularity, divergence on the pilot) and appropriately deferred where it is not (ray-tracing, powered corpus). Safe to seal.