measure(H_1040): 🟢 BASELINE-REGIME-SPECIFIC — pre-rollout-latent uniquely makes big-Φ↓ (H_1033 residual pinned)

[dancinlife]

H_1040 — which baseline regime predicts the big-Φ-DOWN half? (H_1033 residual)

Measures the pre-registered H_1040 (pre-reg merged #1939). 🟢 BASELINE-REGIME-SPECIFIC (H1 PASS).

H_1033 (⚪ degenerate) found the planning split's big-Φ-DOWN half does not reproduce on any
independent-bits baseline (0/5), deferring: the sign is dominated by the BASELINE CONTRAST, not
task structure. H_1040 holds the canonical planning intervention FIXED (planning_trajectories
depth=8 → H_plan, VERBATIM H_973/H_1004) and sweeps 4 baseline regimes.

Per-baseline-regime contrast (planning − baseline, 30 seeds, Cohen d)

baseline regime	big-Φ d	big-Φ ctr	bigΦ-DOWN?	faith d	faith ctr	faith-UP?
independent-bits	+4.221	+3.8089	False	+7.985	+2.8364	True
pre-rollout-latent ★	−1.834	−4.0083	True	+5.178	+2.3332	True
shuffled-time	+3.912	+4.0701	False	+0.000	+0.0000	False
matched-marginal-corr	+2.320	+2.5133	False	+7.885	+2.8055	True

★ = a-priori pre-registered pick. PASS: the pre-rollout-latent baseline ALONE makes big-Φ go
DOWN (d=−1.834 ≤ −0.8) AND faithful go UP (+2.333), while ALL 3 other baselines do NOT make big-Φ
go DOWN → the big-Φ-DOWN half is a planning-vs-(pre-rollout-latent) property, NOT regime-independent.

Key finding

Only the model's OWN prior state (the original H_973 GREEDY, big-Φ=9.53) makes planning LOWER big-Φ.
The big-Φ contrast −4.0083 reproduces the original H_973 number −4.008 exactly → the H_1033
residual is pinned: the big-Φ-DOWN sign is dominated by the BASELINE CONTRAST (the high-Φ prior
state), not by generic task decomposability. faithful is NULL only for time-shuffle (which leaves
the MI matrix unchanged → contrast exactly 0).

Verification (g73 / a_phi_iit4_tool)

• CPU mirror RE-PROVEN ≡ stdlib at n=4 (big-Φ |Δ|=1.34e-10, faithful |Δ|≤3.75e-6) AND n=5
  (|Δ|≤7.97e-10) BEFORE scoring. Phi from stdlib faithful_phi + iit4_bigphi mirrors only — NO proxy.
• SERIAL only (H_1038 Pool-hang lesson). $0 CPU-local, 0 pods/GPU.
• Honest scope: TOY n=4, 4 baselines × 30 seeds; scale-transfer UNVERIFIED. g5 CODE-measured (p7).

Verdict raw: .verdicts/1040_split_baseline_regime/H_1040.txt. Probe: UNIVERSE/h1040_split_baseline_regime.py.

🤖 Generated with Claude Code