Fresh-seed replication of the direction (not the absolute magnitude) of the conditional-MI dissociation reported in O0-STUDY-001

Claim-status banner

CLAIM STATUS         FRESH-SEED REPLICATION SUPPORTED (Tier 2)
                     Direction of dissociation replicates; magnitude
                     inherits STUDY-001's bin-count fragility.
EVIDENCE TYPE        COMPUTATIONAL SIMULATION (same code, fresh seeds)
PHYSICAL VALIDATION  NONE
INDEPENDENT REPLICATION NONE (requires Tier 3: reimplementation)
CONFIRMATORY VS EXPLORATORY  CONFIRMATORY

SOURCE PROPOSITION
  Extends O0-STUDY-001 by testing whether the DIRECTION of its
  conditional-MI finding survives a fresh seed band. The
  hypothesis under test is:
  H_direction: In the same Kuramoto simulation, at bins_z in {4, 6, 8},
  the mean cond_mi_excess for condition C (dense coupling) is
  strictly greater than for condition A (common drive), with
  paired Cohen's d >= 2.0 (median across bin values).

SUPPORTED
  - On seed band 400_000..400_029 (STUDY-001 used 300_000..300_029),
    the direction test passes at all three bin values.
  - Paired Cohen's d (C - A) at bins_z = 4: 3.14
  - Paired Cohen's d (C - A) at bins_z = 6: 2.96
  - Paired Cohen's d (C - A) at bins_z = 8: 2.55
  - Median: 2.96. Sign consistent across all three bins.

NOT ESTABLISHED
  - That the absolute magnitude of cond_mi_excess replicates. As in
    STUDY-001 v1.1.2, the absolute value is bin-count-sensitive and
    the preregistered absolute threshold does not apply here.
  - That the effect survives an INDEPENDENT REIMPLEMENTATION (that
    would be Tier 3, still open).
  - That the effect survives an ALTERNATIVE ESTIMATOR (Tier 4, still
    open).

MOST LIKELY ALTERNATIVE
  Any bug in the shared analysis code (STUDY-001's `analysis.py`)
  would produce a matching false positive here. The Tier-3
  reimplementation is the correct next step to rule out that
  alternative.

NEXT DISCRIMINATING TEST
  Tier 3: independent reimplementation of the Kuramoto simulator
  AND the conditional-MI estimator. If Tier-3 also shows the C > A
  direction at Cohen's d >= 2.0, the replication ladder advances to
  INTERNALLY REPLICATED. If not, the STUDY-001 finding was likely
  an implementation artifact.

Abstract

O/0 Study 001 v1.1.2 concluded that the *direction* of the

C > A dissociation on `cond_mi_excess` (given the mean-field phase

Ψ) is robust across a bin-count sweep, but the *absolute magnitude*

fails the preregistered 0.10-nat threshold at `bins_z >= 8`. The

top-line secondary verdict was INCONCLUSIVE.

We test whether the DIRECTION replicates on a completely fresh seed

band. Using the same Kuramoto simulator, the same

`analysis.condition_result_bins_sweep` function, and the same

baseline config as STUDY-001, we ran 30 confirmatory seeds × 2

conditions (A vs. C) × 3 bin_z values (4, 6, 8).

**Result:** paired Cohen's d (C - A) values [3.14, 2.96, 2.55] across

bin values, median 2.96. Sign consistent across all three bins. The

preregistered SUPPORT rule (median d >= 2.0 AND sign consistent)

passes with margin.

**Verdict:** FRESH-SEED REPLICATION SUPPORTED (Tier 2). This is a

strictly weaker claim than "INTERNALLY REPLICATED," which the

mandate reserves for cases where the replication has been performed

by an independent implementation path. Tier 3 remains open.

Design

• Simulator: `research/studies/O0-STUDY-001/src/simulation.py`
• Analyzer: `research/studies/O0-STUDY-001/src/analysis.py`
• Config: `research/studies/O0-STUDY-001/configs/baseline.json`
• Fresh seed band: 400_000..400_029 (30 seeds)
• Conditions: `A_common_drive`, `C_dense_coupled`
• Bin values: {4, 6, 8}
• Paired Cohen's d computed from seed-paired

`cond_mi_excess` values.

Preregistered SUPPORT rule (frozen)

• Median Cohen's d across bin values >= 2.0, AND
• Sign consistent across all three bin values (all positive means

C > A as predicted).

Actual: median = 2.96, all positive. Passes.

Why this study is worth doing

1. It confirms that STUDY-001's direction finding is not a fluke of

the specific seed band 300_000..300_029.

2. It sets up the necessary Tier-3 replication as the next

discriminating step.

3. It illustrates the utility of separating direction claims from

magnitude claims — a lesson we learned from STUDY-001 v1.1.2's

verdict downgrade.

Files

• `src/run_study.py` — reuses STUDY-001 code
• `results/summary.json` — full per-bin statistics
• No new figure (numbers reported in JSON)

Revision history

• 1.0.0 (2026-07-26) — Initial fresh-seed replication. Direction

test passed; magnitude test intentionally not attempted (known

fragile per STUDY-001 v1.1.2).