Campaign session

01KWY17Z028CZHCKHE1H70267X

finished 2026-07-07 10:16:10.498487+00:00 → 2026-07-07 10:19:45.985812+00:00 · 0 runs · supervisor: react-agent

“BLOCKED — all 3 criteria unresolved due to infra (PostGIS unreachable) + undeclared analysis harness; ~0 CPU-h spent, no science runs. Not a scientific refutation.”

Archive snapshot, as of 11 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.

Success criteria

Criterion	Resolved
Primary (topology): LOS-vs-NLOS mean CSI occupancy-sensitivity difference, 95% CI excluding 0 with \|Cohen's d\| >= 0.2, on >=10 LOS + >=10 NLOS links per floor across >=20 floors	no
Horse-race honesty: floor-level area correlation and link-level topology correlation each reported as CI + effective n (CI comparison, not bare point estimate)	no
Method: real Welch t (Satterthwaite df) + exact permutation p at cluster level + one BH-FDR pass over both topology sessions' full predictor list, m stated	no

Synthesis

Campaign session synthesis — c-csi-topology-los-balanced

Verdict: BLOCKED. All three success criteria UNRESOLVED due to infrastructure unavailability, not scientific refutation. Zero science runs executed; ~0 CPU-h of the 8 CPU-h budget consumed.

A single feasibility probe (exp-csi-static; floor=a_seminar, placement=random, n_agents=4, n_placements=8, fading_seed=0) was fanned out before committing the ~400 balanced-placement runs — a deliberate cheap-evidence-first move so the CPU/token budget would not be spent on doomed runs. It surfaced two independent hard blockers.

Blocker 1 — PostGIS unreachable (fatal). Scene-staging / floor-export aborted in 0.21 s: psycopg.OperationalError: connection to server at '127.0.0.1', port 5435 failed: Connection refused. The GIS database scenegeom queries to build the Sionna scene is not reachable from the runner container. This is floor-independent — every ResPlan+ZInD floor export hits the same DB — so no floor or parameter change recovers it, and a re-run of the same plan is not warranted. The image resolved and the container started, so the curated note's "not-yet-pushed-to-GHCR" caveat was not the blocker.

Blocker 2 — artefact-contract gap (confirmed). Even with PostGIS up, the stock entrypoint emits links.parquet (frame, link_id, n_occluders, mean_amp_db, rician_k_db) + metrics.json — no LOS/NLOS or wall-count columns. The brief's hard-gate topology_linkunit.parquet (per floor×link; cluster bootstrap; Welch + exact permutation; BH-FDR) needs a custom post-processing harness that is not declared (no notebook: block).

Per-criterion status.

C1 (topology LOS−NLOS Cohen's d, 95% CI excl. 0): UNRESOLVED — no balanced-link corpus produced.
C2 (area-vs-topology horse-race CIs): UNRESOLVED — no floor- or link-level estimates exist.
C3 (Welch + exact permutation + BH-FDR method): UNRESOLVED — no data to apply the method to.

Figures: none rendered. figure_count = 0 is correct here (zero corpus → nothing to plot), not an oversight of the mandatory figure request.

Prerequisites before re-dispatch. (1) Make PostGIS reachable from the runner container (bind/route port 5435 on the monad network). (2) Declare a notebook: block or extend the entrypoint to derive topology_linkunit.parquet with LOS/NLOS + wall-count classification and the cluster-correct statistics the brief specifies. Until both hold, this fair test cannot run and the ambiguous "area beats topology" state from c-csi-topology-sniffer-count remains un-retired — neither confirmed nor refuted.

Attached runs

The attached runs are not in the current atlas snapshot — rebuild via web --build-sim-atlas.