c-csi-cross-geometry-scaleout — session 01KWWHJCTHAD3MPTBZD41PQTQG (partial: launch-blocked)
Outcome: 0 runs. All three success criteria UNRESOLVED — blocked before fan-out by an under-specified brief.
The blocker (structural, not budget)
The design requires a nested grid: ~40 floors (random effect) × seeds [0,1,2] × ~5 occupancy levels ≈ 600 runs. But the authoritative sim_params.exp-csi-static block declares only seeds: [0,1,2] — no enumerated floors: axis, no n_agents occupancy grid, no base floor/placement/radio config. The 40-floor set lives solely in prose (c-csi-cross-geometry-resplan + ZInD). Under IP-085 the frontmatter is authoritative and prose is sweep-narrative only; inventing 40 PostGIS floor names to build the Cartesian grid would violate that contract and produce an unauditable corpus. This is the analyser-knob-missing case: the supervisor's sanctioned response is to add the values to the brief's sim_params: block via the brief-design loop, not mid-session.
What was validated
- Powered-n confirmed. n₀ = z²σ²/d² = 1.96²·4.82²/1.5² ≈ 39.7 → ≈40 floors (α=0.05, σ=4.82 from the ResPlan jackknife audit, d=1.5). The brief's "≈40 floors" is correct; the prior 6-floor/single-seed evidence base is underpowered by ~6.6× and touches the decision gate — exactly the external-validity gap this campaign exists to close.
- Plan pre-registered. Primary estimand = LOFO cross-floor inflation ratio (held-out MAE ÷ held-out within-floor leave-one-occupancy-out MAE — Criterion 3's honesty requirement, computable via the
lofo_cvreduction withgroup_metric=floor), bootstrap-over-floors 95% CI, unit of replication = floor (cluster per thompson2012_e41f ch.12). Variance split = nested σ²_floor vs σ²_seed with the geometry-dominates test as declared. Single pre-declared primary; per-floor CV(N) table descriptive (no stars); any band×metric grid gets BH-FDR first. - Determinism caveat flagged. Ray tracing is deterministic given scene+params; only
radio.fading_seedvaries across [0,1,2]. σ²_seed will capture fading-draw variance only — the reduction must confirm seeds are not byte-identical (multi_seed_ciinertflag) or σ²_seed collapses to ~0 and the geometry-dominates claim becomes trivially true for the wrong reason.
Required fix before re-dispatch
Amend experiments/campaigns/c-csi-cross-geometry-scaleout.md frontmatter sim_params.exp-csi-static with (a) an enumerated floors: [...] axis (~40 real PostGIS floor names from ResPlan + ZInD), (b) an occupancy n_agents: [...] grid (~5 levels), (c) a base: with placement, n_placements, radio.*, body.body_loss_db (cite exp-csi-calibration, don't hand-pick). Then dispatch via /campaign-systematic (budget_factor=1.0; ~600 runs, 8–12 CPU-h) as the brief itself prescribes. The image must first be pushed to GHCR (currently local-only; where=remote-vm cannot resolve a digest).
Recommended figures (once corpus exists)
csi_cross_geometry— per-floor CV(N) small multiples + a CI-bearing cross-floor-spread summary panel (the uncertainty view whose absence hid the original finding).seed_trace— per floor-group replication trace (every seed a light trace, mean bold).