Campaign session

01KX3KMP930CASHHAGJR88VHFP

finished 2026-07-09 14:13:54.083546+00:00 → 2026-07-09 14:17:47.210583+00:00 · 1 run · supervisor: react-agent

“Prereq #2 CLEARED (PostGIS 13→113 floors; Phase-2 ResPlan import landed since 12:49). Probe run on newly-imported floor resplan-10156 passed the gate end-to-end (where=local, gis export-floor, 7 artefacts, 5s). C1 now UNBLOCKED but unrun — the powered ~600–1650-run sweep exceeds this session's remaining LLM budget; dispatch a fresh /campaign-systematic run. C2/C3 stand resolved from 01KWYBE.”

Archive snapshot, as of 23 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: the bootstrap-over-floors 95% CI on the LOFO inflation ratio (held-out MAE / within-floor leave-one-occupancy-out MAE) has a lower bound exceeding 1.5x.	no
Variance split: nested decomposition reports sigma^2_floor and sigma^2_seed; the geometry-dominates claim holds iff between-floor SD of per-floor CV(N) exceeds within-floor between-seed SD.	yes
Comparator honesty: the LOFO denominator is the held-out within-floor baseline (leave-one-occupancy-out), never in-sample training residual.	yes

Synthesis

c-csi-cross-geometry-scaleout — session 01KX3KMP930CASHHAGJR88VHFP (2026-07-09): Phase-2 UNBLOCKED, machinery validated, 1 probe run

Verdict: the floor-corpus blocker that sealed the 12:49 session is cleared, and the run chain is validated on the new corpus — but the powered sweep is handed to a fresh dispatch on budget grounds.

What changed since the 12:49 blocked session (01KX3ES7)

Prereq #2 (>=40-floor Phase-2 import) — NOW CLEARED. At 12:49 gis_stats reported 13 floors / 12 experiments (only 6 of 10 Phase-1 floors staged). This session gis_stats reports 113 floors / 116 experiments / 467 devices / 475 experiment_devices — the ResPlan scale-out import batch landed in the interim. n=110 replications >> n0~=40 (z^2 sigma^2/d^2, sigma~=4.82, d=1.5), so C1 is in principle powered.
Prereq #1 (PostGIS reachable) — CLEARED (was already clear at 12:49).

Preflight probe (Step 2.5, 1 run)

One cheap probe (exp-csi-static, floor resplan-10156-floor-0 — explicitly absent at 12:49, so it tests the new import — n_agents=0, n_placements=4, seed=0, 2.4 GHz, where=local): gate_passed=true, exit_code=0, 7 artefacts, 5.0 s wall-clock, run 01KX3KRDC75RMP98HJB0G0FE35. The newly-imported floors stage end-to-end via gis export-floor and the Tx/Rx CSI links resolve. At the probe rate a full 1650-run sweep is ~13 CPU-h (< 24 CPU-h budget) — CPU-feasible.

Criteria status

C1 (primary inflation-ratio CI) — UNRESOLVED, now UNBLOCKED. The corpus is staged and the machinery is validated, but the powered sweep (110 floors x 5 levels x 3 seeds ~= 1650 runs; a 40-floor powered subset ~= 600 runs) was not launched: this session's remaining LLM budget cannot safely poll-to-completion + reduce + synthesize that corpus without risking a mid-flight orphan (the 01KWZC9E partial-grid token-bonfire failure mode). 01KWYBE's under-powered result (pooled x3.37; floor-cluster 95% CI [1.04, 9.85], lower bound 1.04 < 1.5) stands as the last word until the powered run.
C2 (variance split) — RESOLVED (carry-over, 01KWYBEJQK…). sigma^2_floor 0.0109 vs sigma^2_seed 0.000165 (66x); between-floor SD 0.105 vs between-seed 0.0116 (9x). Geometry dominates seed decisively. No new evidence this session.
C3 (comparator honesty) — RESOLVED (carry-over, 01KWYBE…). Held-out leave-one-occupancy-out MAE 0.088 vs within-floor 0.026; in-sample-residual bug fixed. No new evidence this session.

Next action for the operator

Re-dispatch via /campaign-systematic c-csi-cross-geometry-scaleout now. The blocker is gone; a fresh full-budget supervisor (150k tokens, no context-read baggage) can launch the powered Phase-2 sweep. Recommended shape: sim_sweep_launch once per replication (grid n_agents=[0,2,4,6,8] x 2.4 GHz, seeds [0,1,2]), start with a >=40-floor subset to clear C1's gate cheaply, poll via sim_tasks_list(wait_seconds=...), reduce with lofo_cv (inflation ratio + bootstrap-over-floors CI) + nested variance components, then render the mandatory csi_cross_geometry small-multiples + CI panel and seed_trace figures. No figures rendered here (probe emitted empty metrics; nothing to plot).

Honest scope

In-silico, ResPlan geometries, single band (2.4 GHz), CSI-only. A budget-disciplined unblock-and-validate pass: it moves the campaign from "blocked pending import" to "ready for powered dispatch" and confirms the new corpus runs, but does not itself resolve the primary criterion.

Attached runs

Run	Gate	Purpose	Replay
B0G0FE35	✓	—	—