Campaign

Topology vs area for CSI occupancy — a fair fight with LOS/NLOS-balanced placement

c-csi-topology-los-balanced · exp-csi-static

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	Synthesis	Criticism	Figures	Verdict
G3PDM5ZT 2026-07-07T23:20	finished	✓	—	—	Partial seal, 0 runs: feasibility probe re-confirmed all 3 blockers (no stratified-LOS placement key, PostGIS down, corpus <20 floors); campaign not re-dispatchable per brief. No hypothesis retired.
1H70267X 2026-07-07T10:16	finished	✓	—	—	BLOCKED — all 3 criteria unresolved due to infra (PostGIS unreachable) + undeclared analysis harness; ~0 CPU-h spent, no science runs. Not a scientific refutation.
RZG3K6QS 2026-07-06T20:30	finished	✓	—	—	BLOCKED (infra): no runs executed — PostGIS floor-export unreachable on :5435; all three criteria unresolved, zero CPU/S3 consumed.

Brief

Why this campaign exists

c-csi-topology-sniffer-count tried to test whether wall-count-on-path (topology) predicts CSI occupancy sensitivity better than floor area — but the sampled corpus gave topology no chance: 28/40 rows at wall_count=1, 4/40 LOS links. The 2026-07-06 audit's modal lesson applies: more seeds never fix a biased design (peck2008_2ba0 ch. 2). The "area beats topology / topology refuted" memory currently rests on that unfair race and must not drive a hypothesis transition until this fair test runs.

Design

Placement-policy sampling design (not paired arms): each floor's Tx/Rx placements are chosen to yield ≥10 LOS + ≥10 NLOS links per floor, stratified on wall-count-on-path, across ≥20 ResPlan+ZInD floors (floors are what add information for the floor-invariant area covariate; the link-level d ≥ 0.2 criterion needs ~392 links/class, met across floors). ~400 runs, 5–8 CPU-h — dispatch via /campaign-systematic on CI, budget_factor=1.0.

Artefact contract (hard gate)

topology_linkunit.parquet (per floor × link, LOS/NLOS + wall-count columns), predictor correlation table with cluster-correct bootstrap CIs (floors for area-type, links-within-floor blocks for link-level), Welch + exact permutation p, all pushed to S3. Bootstrap numpy-only (IP-087 runner constraint).

Figure render request (mandatory)

LOS/NLOS sensitivity distributions with effective n annotated per class.
seed_trace for the per-floor sensitivity spread.

Honest scope

Non-random floor selection (constructed for LOS/NLOS balance) and non-random covariate assignment → in-silico, non-causal, hypothesis-generation. Either topology earns its place as a predictor or it is cleanly dropped — both outcomes retire the current ambiguous state.

Status — 2026-07-07 (after first dispatch, session 01KWWJ0REMMQ995BQRRZG3K6QS, sealed partial, 0 runs)

Blocked (design surface + infra), correctly caught by the supervisor's feasibility probe before any batch spend:

Design-surface gap (primary): exp-csi-static exposes no config key for LOS/NLOS-balanced Tx/Rx placement stratified on wall-count — the brief's "≥10 LOS + ≥10 NLOS links per floor by construction" is not expressible today. Unblock: a placement-policy mechanism (either a first-class placement: stratified-los config key in the runner, or a host-side scene-staging step that computes wall-count-on-path and places links before dispatch). Until then this campaign must not be re-dispatched.
Infra (shared with all exp-csi-static CI runs): PostGIS on the runner env refused connections on :5435 (2026-07-06); scene staging via gis export-floor is impossible until the service is restored.
Corpus: only 10 ResPlan floors staged in PostGIS (target ≥20) — same import prerequisite as c-csi-cross-geometry-scaleout Phase 2.