c-feature-type-transfer — session 01KWQAVE3C7J8BG57CF6J5A01D synthesis

Verdict: Hypothesis 6 supported in-silico — the cross-environment break is feature TYPE, not RF modality.

Criterion 1 (staging + artefacts): met, with one platform fix. 8 coupled exp-csi-crowd runs across two batches, 4 furnished states each (IP-101 displacement_schedule, Δlayout 0→1.05 m, identical walk seed per batch); all gates passed, links.parquet + ble_links.parquet + trajectory.parquet complete, no NaN/Inf. scene_overrides had to be added to CoupledParamsSchema — first composition of furniture injection with the coupled chain (runs 01KWQB21…, 01KWQB5F…, 01KWQB7K…, 01KWQB9S…; batch 2: 01KWQBKZ…, 01KWQBP4…, 01KWQBR8…, 01KWQBTD…).

Criterion 2 (reduction): met. csi_feature_type_transfer ran standalone host-side (papermill container cannot reach Hetzner S3); its outputs — transfer_curve.parquet with all three families, verdict metrics, figure, and schema — are persisted under the session prefix at s3://monad-knowledge/campaigns/c-feature-type-transfer/sessions/01KWQAVE3C7J8BG57CF6J5A01D/reduction/.

Criterion 3 (feature-type split): met — on batch 2. Batch 1 (IP-101 default 0.6 m wooden boxes, constant-occupancy default crowd) was non-discriminative: furniture perturbed the channel by only 0.004–0.05 dB mean against a 5 dB body-motion signal, vanishing entirely beyond Δ=0.7 m, and constant occupancy (12) left no dynamic range — a platform calibration finding for all layout-drift campaigns. Batch 2 (wardrobe-scale 1.2×0.8×1.8 m wood/metal scatterers, two-wave crowd, occupancy 0→16) produced the split, with a resolved break-point structure: csi-fingerprint breaks at the first displacement (held-out MAE 0.74→22.7 at Δ=0.35 m, staying ~21–23 thereafter), ble-fingerprint holds through Δ=0.35 m (1.87, at its 1.99 control level) and breaks from Δ=0.70 m (10.8) — roughly 2× more layout-robust than the CSI map, consistent with a coarser spatial fingerprint — while csi-kinematic stays flat across all states (0.68→0.83, worst 1.03). The kinematic/fingerprint degradation ratio lands an order of magnitude under the 0.5 criterion (≈0.01); its precise value is estimator-dominated (the csi-fingerprint MAE of ~22 persons on a 0–16 crowd is ridge-extrapolation blow-up), so the load-bearing result is the family ordering: both fingerprints end broken, kinematics do not — a feature-type split, not a modality split.

Criterion 4 (framing): met. In-silico on a self-authored crowd + synthetic CSI/BLE; magnitudes are scene- and estimator-specific and circular against the authored noise. The deliverable is the mechanism — which feature TYPE survives furniture rearrangement. This upgrades Hypothesis 6 (Research Synthesis 2026-07-04) to supported in-silico, reconciling billah2021 vs wang2015/jiang2018 as claims about different feature classes; the D2xx cross-room protocol on real hardware remains the field test, and the real testbed should log kinematic features as the drift-robust channel.

Scope. The corpus is n=1 per cell: one floor (resplan-12439-floor-0), one walk seed per batch, displacement seed 0, one realization per state, four Δ points — no seed replication, so the split's seed-fragility is unmeasured (a live risk given the T* seed-fragility precedent from c-ble-drift-trigger). The BLE break point is bracketed to (0.35, 0.70] m but unresolved within it.

Session notes. CPU ≈ 9 min of the 1.5 h budget (CLI-launched runs; session CPU accounting shows 0 — runs attached via sim_campaign_attach_run). Follow-on candidates: seed-replication sweep (3 walk seeds × 3 displacement seeds) to bound the split's variance; half-step displacement (0.15 m) to locate the BLE-fingerprint break point; a python-ml domain-adversarial arm (H6 next-step 2).

Campaign Critic — session 01KWQAVE3C7J8BG57CF6J5A01D

1. Claim audit

• C1 (staging + artefacts): supported. 8 runs, all gate_passed, links/ble_links/trajectory.parquet present, no NaN/Inf. Batch-2 trajectory.parquet is byte-identical (93 961 B) across state 0 and state 3 — corroborates identical-walk, furniture-only delta. Note: 8 runs (2 batches), not the 4-state staging the criterion literally specifies; batch-2 satisfies intent.
• C2 (reduction + transfer_curve.parquet): borderline at review time. The reduction ran host-side and its outputs were not initially in S3. (Resolved before close: outputs persisted under the session reduction/ prefix.)
• C3 (feature-type split): supported, with caveat. Crowd design confirmed (two-wave, occupancy 0→16; batch-1 static-12). The ordering — two CSI families at opposite extremes, BLE between — is a genuine type-split, not modality. But the 0.011 ratio is dominated by ridge extrapolation blow-up (csi-fingerprint MAE ~22 on a 0–16 crowd is nonphysical).
• C4 (framing): supported. Disciplined; mechanism-not-magnitude, feeds D2xx.

2. Weak claims

• "Degradation ratio 0.011" quoted to 3 figures — driven by an unstable estimator, not a robust magnitude. Report the ordering.
• "Both fingerprints break" flattens a real gap: ble-fingerprint is flat/better through Δ=0.35 (1.87 ≤ baseline 1.99) and only breaks at Δ≥0.70 — ~2× more layout-robust than csi-fingerprint.
• C2 "met" — the load-bearing artefact was not in the corpus at review time.

3. Honest scope

The corpus is n=1: one floor (resplan-12439-floor-0), one walk seed per batch, displacement seed 0, one realization per state, four delta points — no error bars, no seed replication, so nothing can be said about the split's seed-fragility (a live risk per prior T* work). The channel is self-authored (−30.9 dB/person atten, fixed RSSI slope); the fingerprint collapse may partly reflect ridge extrapolation instability rather than pure spatial-map decorrelation. The BLE break point (0.35→0.70) is unresolved. Does not settle billah2021 vs wang2015/jiang2018 against hardware.

Severity: medium. Agrees with synthesis verdict: yes. All three weak claims were integrated into the sealed synthesis (ratio de-emphasised to ordering, BLE break-point structure surfaced, reduction outputs persisted to the session prefix).