Sim-to-real count transfer — session 01KXBTWW2CM855Z7CW1W5RD4M5 (PARTIAL / BLOCKED)

Executive summary

This session opened, ran the mandated preflight (Step 2.5), and sealed clean with zero runs launched. The campaign is blocked on two prerequisites that no parameter change or CPU spend can recover this session. Sealing a confirmed block for ~$1 is the correct outcome; launching the 24-cell sionna sweep would have orphaned an unsupervisable in-flight corpus.

What the preflight established

1. No cheap schema check exists for the CSI family. sim_scenario_validate(target="params") validates only the pedestrian-walk ParamsSchema; it rejected experiment, radio, doppler, body, trajectory_frame_stride as extra_forbidden. The exp-csi-crowd (sionna) params are therefore unverifiable through the MCP authoring-loop validator. Confirming the emitted-artefact contract requires a live jupedsim->sionna probe run.
2. The impairment channel is undeclared. The brief's sim_params.exp-csi-crowd.base carries only doppler.mode: cfr — there is no impairment (CFO/STO/phase-noise, IP-110) configuration block. Criterion 1 hard-fails the session if csi_impaired.hdf5 is absent ("the dominant sim-to-real gap is hardware artefacts, guarino2026 / chen2023"). Whether the current runner emits that channel from these params is unconfirmed and unconfigured.
3. The payoff reduction does not exist. The brief itself states csi_sim2real_count_transfer.py "is authored after the synthetic corpus lands (it needs the run S3 paths)" and is "dispatchable via csi-reduction.yml once written." Criteria 2 and 3 are computed only by this reduction. No fan-out of synthetic runs can satisfy them until it is written.

Per-criterion result

• Criterion 1 (corpus artefact contract): UNRESOLVED. Impairment-channel params are not declared in the brief and the emission of csi_impaired.hdf5 cannot be confirmed without a live probe. Needs an explicit IP-110 impairment config in sim_params.base (or confirmation the runner injects it), verified by one cheap probe run.
• Criterion 2 (transfer test): UNRESOLVED — structurally. Requires csi_sim2real_count_transfer.py, which is unwritten. Cannot be produced by any run this session.
• Criterion 3 (honesty gate): UNRESOLVED — structurally. Same reduction; also requires the zero-shot floor + clean-vs-impaired ablation, all downstream of the same unwritten harness. Note the real-only Doppler baseline to beat is LOEO R2~=0.47 (established 2026-07-12).

Prerequisites to unblock (ordered)

1. Author csi_sim2real_count_transfer.py (consumes this campaign's csi_impaired.hdf5 + streams WiMANS; LOEO, bootstrap CIs per k in {5,10,20}); register it for csi-reduction.yml.
2. Add an explicit IP-110 impairment block to sim_params.exp-csi-crowd.base (CFO/STO/phase-noise) OR document that exp-csi-crowd emits csi_impaired.hdf5 by default; then run ONE cheap probe cell (n_agents=1, seed=0, where=remote-vm) to confirm csi.hdf5 + csi_impaired.hdf5 + links.parquet all emit with no NaN/Inf.
3. Only then dispatch the 24-cell corpus sweep (grid scenario.n_agents in {0..5} x seeds {0..3} at 5 GHz) — well inside the 12 CPU-h budget — and run the transfer reduction.

Statistics note

No quantitative claim is made — no corpus was produced, so no estimate/interval/effect-size is reportable. The declared seed axis (4 seeds x 6 densities) is a reasonable base for the corpus but must be power-checked against the transfer effect size once the reduction defines the estimand (per-k LOEO R2/MAE, unit of replication = environment fold, paired sim-pretrain-vs-real-only within fold).

Real-life meaning

The scientific question — can synthetic-impaired CSI pretraining beat real-only at a small real budget — remains open and worth pursuing; the meta-hypothesis is licensed now that a validated real anchor exists (WiMANS LOEO R2~=0.47 in Doppler). This session simply confirms the campaign is not yet executable: the analysis harness and the impairment-channel contract must land first. This is a wiring gap, not a negative scientific result.